Numéro de publication | US7388954 B2 |

Type de publication | Octroi |

Numéro de demande | US 10/178,560 |

Date de publication | 17 juin 2008 |

Date de dépôt | 24 juin 2002 |

Date de priorité | 24 juin 2002 |

État de paiement des frais | Payé |

Autre référence de publication | CN1672341A, CN1672341B, EP1516437A1, US20030235312, WO2004002002A1 |

Numéro de publication | 10178560, 178560, US 7388954 B2, US 7388954B2, US-B2-7388954, US7388954 B2, US7388954B2 |

Inventeurs | Lucio F. C. Pessoa, Roman A. Dyba, Perry P. He |

Cessionnaire d'origine | Freescale Semiconductor, Inc. |

Exporter la citation | BiBTeX, EndNote, RefMan |

Citations de brevets (67), Citations hors brevets (12), Référencé par (8), Classifications (6), Événements juridiques (14) | |

Liens externes: USPTO, Cession USPTO, Espacenet | |

US 7388954 B2

Résumé

A communication system having an echo canceller is disclosed. One embodiment of the echo canceller includes an adaptive filter used to provide an estimate of reflected echo which is removed from the send signal. The echo canceller may also include a near-end talker signal detector which may be used to prevent the adaptive filter from adapting when a near-end talker signal is present. The echo canceller may also include a nonlinear processor used to further reduce any residual echo and to preserve background noise. The echo canceller may also include a monitor and control unit which may be used to monitor the filter coefficients and gain of the adaptive filter to maintain stability of the echo canceller, estimate pure delay, detect a tone, and inject a training signal. The echo canceller may also include a nonadaptive filter used to reduce the length of the adaptive filter.

Revendications(18)

1. A method for indicating whether a tone is present in a communication signal, comprising:

delaying the communication signal by a predetermined first delay to produce a first delayed signal;

delaying the communication signal by a predetermined second delay to produce a second delayed signal;

performing a polynomial function using the first delayed signal, the second delayed signal, and the communication signal to produce an estimate signal, wherein performing the polynomial function comprises:

multiplying the first delayed signal by the first delayed signal to produce a first product;

multiplying the second delayed signal by the communication signal to produce a second product;

subtracting the second product from the first product to produce an intermediate signal;

taking a magnitude of the intermediate signal to produce the estimate signal;

low-pass filtering the estimate signal to produce a smooth estimate signal; and

using the smooth estimate signal to indicate whether the tone is present.

2. A method as in claim 1 , wherein the smooth estimate signal is a power estimate signal.

3. A method as in claim 1 , further comprising:

taking a magnitude of an intermediate signal to produce a magnitude signal.

4. A method as in claim 1 , further comprising:

using an impulsive change of the smooth estimate signal to detect a phase change of the communication signal.

5. A method for indicating whether a tone is present in a communication signal, comprising:

delaying the communication signal by a predetermined first delay to produce a first delayed signal;

delaying the communication signal by a predetermined second delay to produce a second delayed signal;

performing a polynomial function using the first delayed signal, the second delayed signal, and the communication signal to produce an estimate signal;

low-pass filtering the estimate signal to produce a smooth estimate signal;

delaying the smooth estimate signal by a predetermined third delay to produce a delayed smooth estimate signal;

determining a minimum estimate between the smooth estimate signal and the delayed smooth estimate signal; and

determining a maximum estimate between the smooth estimate signal and the delayed smooth estimate signal.

6. A method as in claim 5 , wherein the step of using the smooth estimate signal to indicate whether the tone is present further comprises:

computing a ratio of the minimum estimate and the maximum estimate;

comparing the ratio to a predetermined ratio threshold;

repeating the step of computing the ratio;

determining if the ratio has continuously exceeded the predetermined ratio threshold for a sufficient number of steps of repeating; and

if the ratio has continuously exceeded the predetermined ratio threshold for the sufficient number of steps of repeating, providing a tone indicator signal which indicates that the tone has been detected.

7. A method as in claim 6 , wherein the step of using the smooth estimate signal to indicate whether the tone is present further comprises:

determining if the smooth estimate signal has exceeded a predetermined ratio threshold for a sufficient number of steps of repeating.

8. A method as in claim 6 , wherein the smooth estimate signal is a power estimate signal and wherein the step of using the smooth estimate signal to indicate whether the tone is present further comprises:

determining if the power estimate signal has exceeded a predetermined low power threshold; and

if the power estimate signal has not exceeded a predetermined low power threshold, waiting until the power estimate signal has exceeded the predetermined low power threshold before performing the steps of computing, comparing, repeating, and determining if the ratio has continuously exceeded the predetermined ratio threshold.

9. A method for indicating whether a tone is present in a communication signal, comprising:

delaying the communication signal by a predetermined first delay to produce a first delayed signal;

delaying the communication signal by a predetermined second delay to produce a second delayed signal;

performing a polynomial function using the first delayed signal, the second delayed signal, and the communication signal to produce an estimate signal;

low-pass filtering the estimate signal to produce a smooth estimate signal;

using the smooth estimate signal to indicate whether the tone is present;

delaying the communication signal by a predetermined third delay to produce a third delayed signal;

generating a generated sinusoidal signal with the predetermined single frequency;

multiplying the generated sinusoidal signal by the communication signal to produce a correlation signal;

multiplying the generated sinusoidal signal by the third delayed signal to produce a delayed correlation signal;

low-pass filtering the correlation signal to produce a smooth correlation estimate;

low-pass filtering the delayed correlation signal to produce a delayed smooth correlation signal;

using the smooth correlation signal and the delayed smooth correlation signal to detect whether the tone is the predetermined single frequency tone; and

detecting phase reversal of the predetermined single frequency tone.

10. A method for indicating whether a tone is present in a communication signal, comprising:

delaying the communication signal by a predetermined delay to produce a delayed signal;

generating a generated sinusoidal signal;

performing a polynomial function using the delayed signal, the generated sinusoidal signal, and the communication signal to produce a first correlation signal and a second correlation signal, wherein performing the polynomial function comprises:

multiplying the generated sinusoidal signal by the communication signal to produce the first correlation signal; and

multiplying the generated sinusoidal signal by the delayed signal to produce the second correlation signal; and

using the first correlation signal and the second correlation signal to indicate whether the tone has been detected.

11. A method as in claim 10 , wherein the tone is a predetermined single frequency tone.

12. A method for indicating whether a tone is present in a communication signal, comprising:

delaying the communication signal by a predetermined delay to produce a delayed signal;

generating a generated sinusoidal signal;

performing a polynomial function using the delayed signal, the generated sinusoidal signal, and the communication signal to produce a first correlation signal and a second correlation signal;

low-pass filtering the first correlation signal to produce a smooth correlation estimate;

low-pass filtering the second correlation signal to produce a delayed smooth correlation signal; and

using the first correlation signal and the second correlation signal to indicate whether the tone has been detected, wherein using comprises:

determining a magnitude of the smooth correlation estimate;

determining a magnitude of the delayed smooth correlation estimate; and

determining an effective smooth correlation signal by selecting a maximum between the magnitude of the smooth correlation estimate and the delayed smooth correlation estimate.

13. A method as in claim 12 wherein the step of using the first correlation signal and the second correlation signal to indicate whether the tone is present further comprises:

delaying the effective smooth correlation signal by a second predetermined delay to produce a delayed effective smooth correlation signal;

determining a minimum estimate between the effective smooth correlation signal and the delayed effective smooth correlation signal; and

determining a maximum estimate between the effective smooth correlation signal and the delayed effective smooth correlation signal.

14. A method as in claim 13 , wherein the step of using the first correlation signal and the second correlation signal to indicate whether the tone is present further comprises:

computing a ratio of the minimum estimate and the maximum estimate;

comparing the ratio to a predetermined ratio threshold;

repeating the step of computing the ratio;

determining if the ratio has continuously exceeded the predetermined ratio threshold for a sufficient number of steps of repeating;

if the ratio has continuously exceeded the predetermined ratio threshold for the sufficient number of steps of repeating, providing a tone detector signal which indicates that the tone has been detected.

15. A method for indicating whether a tone is present in a communication signal, comprising:

delaying the communication signal by a predetermined delay to produce a delayed signal;

generating a generated sinusoidal signal;

performing a polynomial function using the delayed signal, the generated sinusoidal signal, and the communication signal to produce a first correlation signal and a second correlation signal;

using the first correlation signal and the second correlation signal to indicate whether the tone has been detected;

if the tone has been detected, detecting a phase sign of the tone; and

wherein the step of detecting the phase sign of the tone comprises determining a sign of one of the first and second correlation signals whose magnitude is equal to the effective smooth correlation signal.

16. A method as in claim 15 , further comprising:

if the tone is detected, determining if a phase reversal of the tone occurs by monitoring the sign of the one of the first and second correlation signals whose magnitude is equal to the effective smooth correlation signal.

17. A method as in claim 16 , further comprising:

if the phase reversal of the tone occurs, counting a number of times the phase reversal of the tone occurs.

18. A tone detector for detecting a tone in a communication signal, comprising:

a first circuit for providing a first delayed signal;

a second circuit for providing a second delayed signal;

estimator circuitry for performing a polynomial function using the first delayed signal, the second delayed signal, and the communication signal to produce an estimate signal, wherein the estimator circuitry comprises;

a first multiplier for multiplying the first delayed signal by the first delayed signal to produce a first product;

a second multiplier for multiplying the second delayed signal by the communication signal to produce a second product;

a subtractor for subtracting the second product from the first product to produce an intermediate signal; and

magnitude circuitry for taking a magnitude of the intermediate signal to produce the estimate signal;

a low-pass filter for filtering the estimate signal to provide a smooth estimate signal; and

tone indication circuitry for using the smooth estimate signal to indicate whether the tone is present.

Description

This is related to U.S. patent application having attorney docket number SC12026TS, filed on even date, and entitled “Monitoring and control of an Adaptive Filter in a Communication System,” U.S. patent application having attorney docket number SC12120TS, filed on even date, and entitled “Method and Apparatus for Pure Delay Estimation in a Communication System,” U.S. patent application having attorney docket number SC11993TS, filed on even date, and entitled “Method and Apparatus for Tone Indication,” and U.S. patent application having attorney docket number SC12107TS, filed on even date, and entitled “Method and Apparatus for Performing Adaptive Filtering,” all of which are assigned to the current assignee hereof.

The present invention relates generally to non-linear processing, and more specifically, to a method and apparatus for non-linear processing of an audio signal.

Echo cancellation is used in a telecommunication network (such as in a Public Switching Telephone Network (PSTN) or Packet Telephony (PT) network) to ensure voice quality through elimination or reduction of electric or line echo from the telecommunication network. The source of this electric or line echo may be the impedance mismatch of a hybrid circuit which is a device used to convert signals from a four-wire communication network interface to a two-wire local subscriber loop and vice versa. Echoes with long delays in the communication network may be noticeable which may create significant or even unbearable disturbance during telephone voice communication. Therefore, a need exists for an echo canceller that is able to eliminate the echoes completely or to reduce them to an acceptable level within the telecommunication network. Also, a need exists for an echo canceller that is capable of detecting tones received via the telecommunication network while maintaining stability.

The present invention is illustrated by way of example and not limited by the accompanying figures, in which like references indicate similar elements, and in which:

Skilled artisans appreciate that elements in the figures are illustrated for simplicity and clarity and have not necessarily been drawn to scale. For example, the dimensions of some of the elements in the figures may be exaggerated relative to other elements to help improve the understanding of the embodiments of the present invention.

As used herein, the term “bus” is used to refer to a plurality of signals or conductors which may be used to transfer one or more various types of information, such as data, addresses, control, or status. The conductors as discussed herein may be illustrated or described in reference to being a single conductor, a plurality of conductors, unidirectional conductors, or bidirectional conductors. However, different embodiments may vary the implementation of the conductors. For example, separate unidirectional conductors may be used rather than bidirectional conductors and vice versa. Also, plurality of conductors may be replaced with a single conductor that transfers multiple signals serially or in a time multiplexed manner. Likewise, single conductors carrying multiple signals may be separated out into various different conductors carrying subsets of these signals. Therefore, many options exist for transferring signals.

The terms “assert” and “negate” (or “deassert”) are used when referring to the rendering of a signal, status bit, or similar apparatus into its logically true or logically false state, respectively. If the logically true state is a logic level one, the logically false state is a logic level zero. And if the logically true state is a logic level zero, the logically false state is a logic level one. The symbols “*” and “.” both indicate a multiplication operation. A FIFO or other type of data storage may be used to provide the delays used throughout this invention document.

Also, note that in the descriptions herein, variable names are generally used consistently with each group of related figures. Some variable names, though, may be reused to refer to different things in different groups of related figures. For example, in reference to a particular group of figures, M may refer to a measurement cycle, and in reference to a different group of figures, M may be used as a counter value. The description of each variable name in the equations and figures below, though, will be provided as they are used.

Connectivity

**10**. Communication system **10** includes transmitter/receiver **12**, interface **13**, hybrid circuit **16** (also referred to as hybrid **16**), echo canceller **20**, communication network **24**, echo canceller **22**, interface **15**, hybrid **18**, and transmitter/receiver **14**. Interface **13** includes hybrid **16** and interface **15** includes hybrid **18**. Transmitter/receiver **12** is bidirectionally coupled to hybrid **16** (where, in one embodiment, transmitter/receiver **12** is coupled to hybrid **16** via a two-wire connection such as a twisted pair). Hybrid **16** is coupled to echo canceller **20**, providing a send signal Sin **37** to echo canceller **20** via unidirectional conductors and receiving a receive signal Rout **40** from echo canceller **20** via unidirectional conductors (where, in one embodiment, each of Sin **37** and Rout **40** are provided and received via a wire pair). Echo canceller **20** is coupled to communication network **24** and provides an echo cancelled send signal Sout **42** to communication network **24** and receives Rin **43** from communication network **24**.

Similarly, transmitter/receiver **14** is bidirectionally coupled to hybrid **18** (where, in one embodiment, transmitter/receiver **14** is coupled to hybrid **18** via a two-wire connection such as a twisted pair). Hybrid **18** is coupled to echo canceller **22** via unidirectional conductors for providing signals to echo canceller **22** and unidirectional conductors for receiving signals from echo canceller **22** (where, in one embodiment, each set of unidirectional conductors may be a twisted wire pair). Echo canceller **22** is coupled to communication network **24** and provides an echo cancelled send signal to communication network **24** and receives a received signal from communication network **24**. Control **17** may be a control bus that includes one or more control signals that may be provided to each of transmitter/receiver **12**, hybrid **16**, echo canceller **20**, communication network **24**, echo canceller **22**, hybrid **18**, and transmitter/receiver **14**, as needed. Therefore, in one embodiment, control **17** is coupled to every unit within communication system **10**, while in alternate embodiments, only a portion of the units may require communication with control **17**.

**20** of **20** is referred to as the near end echo canceller while echo canceller **22** is referred to as the far end echo canceller. However, it should be appreciated that the echo canceller illustrated in **22** in the case where echo canceller **22** is at the near end and echo canceller **20** at the far end of communication system **10**.) Echo canceller **20** includes DC notch filter **45**, optional non-adaptive filter **31**, adder **34**, optional non-adaptive filter **35**, gain control **33**, nonlinear processor **32**, near-end signal detector **26**, adaptive filter **28**, monitor and control unit **30**, DC notch filter **49**, and adder **36**. DC notch filter **45** receives Sin **37** and outputs Sin **38** to near-end signal detector **26** and monitor and control unit **30**. If non-adaptive filter **31** is present, then Sin **38** is also provided to non-adaptive filter **31** which is coupled to receive controls from monitor and control unit **30** and outputs Sin **39** to adder **34**. However, if non-adaptive filter **31** is not present, then Sin **38** is the same as Sin **39** which is input to adder **34**. Adder **34** receives Sin **39** and echo estimation signal **48** from adaptive filter **28** and provides an error signal **46** to gain control **33**, near-end signal detector **26**, and monitor and control unit **30**. Gain control **33** is bidirectionally coupled to monitor and control unit **30** and is coupled to provide error signal **47** to nonlinear processor **32**. If non-adaptive filter **35** is present in echo canceller **20**, then, in one embodiment, gain control **33** is within non-adaptive filter **35** which also receives error signal **46**, is bidirectionally coupled to monitor and control unit **30** and provides error signal **47**. Nonlinear processor **32** is bidirectionally coupled to monitor and control unit **30** and provides Sout **42**. Monitor and control unit **30** is also coupled to control **17**, receives Rin **43**, provides training signal **41** to adder **36**, receives Rin **44** from DC notch filter **49**, and is bidirectionally coupled to adaptive filter **28** and near-end signal detector **26**. DC notch filter **49** receives the output of adder **36** (Rout **40**) and provides Rin **44** to near-end signal detector **26**, adaptive filter **28**, and monitor and control unit **30**. Adder **36** receives training signal **41** and Rin **43** and provides Rout **40**.

**26** of **26** includes near-end signal level estimator **50**, far-end signal level estimator **52**, Sin signal level estimator **54**, background processor **56**, near-end signal detection threshold selector **58**, and near-end signal detector **60**. Near-end signal level estimator **50** receives error signal **46** and is coupled to near-end signal detector **60**. Far-end signal level estimator is coupled to receive Rin **44** and is also coupled to near-end signal detection threshold selector **58**. Sin signal level estimator **54** is coupled to receive Sin **38**, and is also coupled to near-end signal detector **60**. Background processor **56** is coupled to monitor and control unit **30**, near-end signal detection threshold selector **58**, and near-end signal detector **60**. Near-end signal detector **60** is also coupled to near-end signal detection threshold selector **58** and monitor and control unit **30**.

**28** of **28** includes adaptive filter **62**, optional non-adaptive filter **64**, and optional delay **66**. Assuming both non-adaptive filter **64** and delay **66** are present in adaptive filter **28**, delay **66** receives Rin **44**, and is coupled to non-adaptive filter **64** and monitor and control unit **30**. Non-adaptive filter **64** is coupled to delay **66**, adaptive filter **62**, and monitor and control unit **30**. Adaptive filter **62** is coupled to receive error signal **46** and coupled to provide echo estimation signal **48**, and is also coupled to monitor and control unit **30**. If non-adaptive filter **64** is not present, then delay **66** is coupled directly to adaptive filter **62**. If delay **66** is not present, then non-adaptive filter **64** receives Rin **44**. If neither delay **66** nor non-adaptive filter **64** are present, adaptive filter **62** receives Rin **44**.

**32** of **32** includes signal level estimator **68**, nonlinear processor controller **74**, and adaptive background level estimator **96**, and is bidirectionally coupled to monitor and control unit **30**. Signal level estimator **68** includes near-end signal level estimator **70** and far-end signal level estimator **72**. Nonlinear processor controller **74** includes nonlinear processor ON controller **76**, nonlinear processor OFF controller **78**, comfort noise generator **86**, noise level matcher **82**, and output signal mixer **84**. Adaptive background level estimator **96** includes short-term background level estimator **88**, background level estimator controller **90**, long-term background level estimator **92**, and background level adapter **94**. Near-end signal level estimator **70** receives error signal **47** and is coupled to nonlinear processor ON controller **76** and background level estimator controller **90**. Far-end signal level estimator **72** receives Rin **44** and is coupled to nonlinear processor ON controller **76**, nonlinear processor OFF controller **78**, and background level estimator controller **90**. Nonlinear processor ON controller **76** and nonlinear processor OFF controller **78** are coupled to noise generator **86** which is coupled to noise level matcher **82**. Output signal mixer **84** is coupled to noise level matcher **82**, receives error signal **47**, and provides Sout **42**. Short-term background level estimator **88** is coupled to background level adapter **94** and receives error signal **47**. Background level estimator controller **90** is coupled to short-term background level estimator **88** and long-term background level estimator **92**. Long-term background level estimator **92** receives error signal **47** and is coupled to background level adapter **94** which is coupled to noise level matcher **82**.

**30** which includes a gain monitor **100** and a filter coefficient monitor **102**. Gain monitor **100** receives Sin **38**, error signal **46**, and is coupled to adaptive filter **28** and gain control **33**. Filter coefficient monitor **102** is coupled to adaptive filter **28**.

**30** which includes decimation filters **104** and **108**, decimators **106** and **110**, near-end signal detector **114**, optional comparator **112**, Echo Return Loss Enhancement (ERLE) estimator **116**, power estimators **120** and **118**, adaptive filter system **128**, and noise generator **132**. Adaptive filter system **128** includes adaptive filter **122**, maximum value locator **124**, and delay determination **126**. Decimation filter **104** receives Rin **44** and is coupled to decimator **106**. Decimation filter **108** receives Sin **38** and is coupled to decimator **110**. Decimator **106** is coupled to near-end signal detector **114**, power estimator **120**, and adaptive filter **122**. Power estimator **120** and near-end signal detector **114** are coupled to adaptive filter system **128**. Optional comparator **112**, if present in monitor and control unit **30**, receives error signal **46** and Sin **38**, and is coupled to adaptive filter system **128**. Decimator **110** is coupled to power estimator **118** and adaptive filter **122**. Power estimator **118** is coupled to ERLE estimator **116** and adaptive filter system **128**, and adaptive filter **122** is coupled to near-end signal detector **114**, ERLE estimator **116**, and maximum value locator **124**. Maximum value locator **124** is coupled to delay determination **126** which provides estimated delay **130** to adaptive filter **28**. Noise generator **132** receives Rin **43** and is coupled to provide injected signal **41** to adder **36**. The portion of monitor and control unit **30** of **17**.

**30** including storage **150**, power estimator **134**, smooth correlator **152**, and tone indication decision unit **166**. Power estimator **134** includes delay **136**, delay **138**, multipliers **140** and **142**, adder **144**, magnitude **146**, and low-pass filter **148**. Smooth correlator **152** includes delay **154**, multipliers **156** and **158**, low-pass filters **160** and **162**, and oscillator **164**. Storage **150** is coupled to delay **136**, delay **138**, low-pass filter **148**, delay **154**, low-pass filters **160** and **162**, and oscillator **164**. Delay **136** receives Rin **44** or Sin **38** and is coupled to delay **138** and multiplier **142**. Delay **138** is coupled to multiplier **140** which also receives Rin **44** or Sin **38**. Adder **144** is coupled to multipliers **140** and **142** and magnitude **146** which is coupled to low-pass filter **148** which is coupled to tone indication decision unit **166**. Delay **154** receives Rin **44** or Sin **38**, and is coupled to multiplier **156**. Multiplier **158** also receives Rin **44** or Sin **38** and is coupled to low-pass filter **160**, oscillator **164**, and multiplier **156**. Multiplier **156** receives delay **154** and is coupled to low-pass filter **162** and oscillator **164**. Tone indication decision unit **166** receives R_{0}(n) from low-pass filter **160** and R_{1}(n) from low-pass filter **162** and provides tone indicator signal **168** to adaptive filter **28**.

Note that **10** and echo canceller **20**. Alternate embodiments may include various different elements than those illustrated, more elements than those illustrated, or less elements than those illustrated, depending on the functionality desired. Furthermore, the blocks within

Operation:

Transmitter/receiver **12**, provides and receives data signals to and from hybrid **16**. Hybrid **16** provides for a four-wire to two-wire conversion between transmitter/receiver **12** and communication network **24**. Therefore, transmitter/receiver **12** can be any device used for communicating over communication network **24**, such as, for example, a telephone or a modem, that is coupled to hybrid **16** via a two-wire subscriber line. Therefore, hybrid **16** provides an interface between a local subscriber loop (having transmitter/receiver **12**) and a communication network (communication network **24**). Transmitter/receiver **14** and hybrid **18** functional analogously to transmitter/receiver **12** and hybrid **16**, respectively.

In communications between transmitter/receiver **12** and transmitter/receiver **14**, electrical or line echo is introduced into the communication by hybrid **16** and hybrid **18**. The source of this echo is the impedance mismatch within hybrid **16**, as well as the impedance mismatch within hybrid **18**. For example, if the impedance within hybrid **16** were perfectly matched, all of the energy from received signal Rout **40** would be transmitted to transceiver/receiver **12**. However, if there is any impedance mismatch within hybrid **16**, some of the energy from received signal Rout **40** would be reflected back through send signal Sin **37**. If the round trip delay through communication network **24** (from transmitter/receiver **14**, in the case of echo introduced by hybrid **16**) is sufficiently long, the reflected echo received by transmitter/receiver **14** from Sin **37** will be noticeable during the communication. This may result in noticeable echoes or even unbearable disturbance during a telephone voice communication. In one example, a sufficiently long delay may refer to a round trip delay of greater than 40 milliseconds. As the round trip delay increases, the echoes may become worse and thus more noticeable and disruptive. (If, on the other hand, the round trip delay is significantly smaller, the echo may not be disruptive since it may be indistinguishable from the side tone.) The round trip delay may include a variety or combination of different delays, including transmission delay, processing delay, computation delay, etc. Depending on the communication system, the round trip delay may be sufficiently large to disrupt communication. Therefore, echo cancellers **20** and **22** may be used to reduce the line echo in communication system **10**. For example, the echo introduced by hybrid **16** from a signal received via Rout **40** (from transmitter/receiver **14**) and reflected back via Sin **37** is processed via echo canceller **20** to reduce the reflected echo prior to sending the signal Sout **42** through communication network **24** back to transmitter/receiver **14**.

As discussed above, line echo is introduced by the impedance mismatch within hybrid **16** and the impedance mismatch within hybrid **18**. Also, acoustic echo may be introduced into the communication via transmitter/receiver **12** and transmitter/receiver **14**. For example, if transmitter/receiver **12** is a speaker phone, the received signal, after being output via the speaker, will bounce around the surrounding environment, and some of the signal may be redirected back into the microphone of transmitter/receiver **12** and also be reflected back to transmitter/receiver **14**. In one embodiment, echo canceller **20** may also function to reduce some aspects of acoustic echo in addition to line echo.

In one embodiment, communication network **24** may include a packet telephony network (including, for example, voice over internet protocol (IP), data over packet, asynchronous transfer mode (ATM), etc., and could either apply to wireless or wireline systems) or Public Switching Telephone Network (PSTN). In alternate embodiments, communication system **10** may refer to any type of communication system. Any communication pathway may be used as interface **13** or interface **15**.

Control **17** provides a control pathway among transmitter/receiver **12** and **14**, hybrid **16** and **17**, echo canceller **20** and **22**, and communication network **24**. Control signals transmitted via control **17** are generally not in-line signals. For example, control **17** may include an enabling/disabling signal to enable or disable echo canceller **20** or **22**. Control **17** may also include a signal to indicate whether the telephone is on or off the hook.

In the embodiments described herein, transmitter/receiver **12** will be referred to as the near end with respect to echo canceller **20** and transmitter/receiver **14** will be referred to as the far end with respect to echo canceller **20**. Therefore, the embodiments herein will be discussed with reference to echo canceller **20**; however, it should be understood that echo canceller **22** operates analogously to echo canceller **20**. That is, in an alternate embodiment, transmitter/receiver **14** may be referred to as the near end with respect to echo canceller **22** and transmitter/receiver **12** the far end with respect to echo canceller **22**.

**20**, where, as mentioned above, transmitter/receiver **12** is the near end and transmitter/receiver **14** is the far end. Sin **37** is the send signal transmitted from transmitter **12**, via hybrid **16**. Echo canceller **20** provides an echo cancelled send signal Sout **42** to receiver **14** via communication network **24** and hybrid **18**. Rin **43** is a receive signal received from transmitter **14** via hybrid **18** and communication network **24**. Echo canceller receives Rin **43** and provides this send signal Rin **43** as Rout **40** to receiver **12** via hybrid **16**.

As discussed above, Sin **37** may include reflected echo introduced by the impedance mismatch within hybrid **16**. Therefore, echo canceller **20** reduces (or eliminates) the introduced reflected echo and provides the echo cancelled send signal Sout **42**. That is, if the impedance in hybrid **16** is perfectly matched, a signal received at the input of the hybrid **16** (e.g. Rout **40**) would result in virtually no response from hybrid **16** (at Sin **37**) because there would be no reflected echo (in the ideal and practically unattainable case). However, if the hybrid is in imbalanced state (a typical case, e.g. where the impedance is mismatched), a signal received via Rout **40** results in a response as shown in **40**) and output (Sin **37**) is illustrated in **28** within echo canceller **20** attempts to “imitate” the hybrid response of Sin **37** (to any input signal Rout **40**) and subtracts it out via adder **34**. Note that the signal Rout **40** is linearly distorted (including its pure transposition in time, i.e., it is shifted in time by a parameter called pure delay). This distortion can be illustrated in the impulse response of the hybrid **16** of **1** in **4**+T**2** in **40**/Sin **37** input/output ports).

Sin **37** is provided to DC notch filter **45** to remove the DC component from Sin **37**. Note that in an alternate embodiment, a high pass filter may be used in place of DC notch filter **45**. Similarly, the output of adder **36** (Rout **40**) is provided to DC notch filter **49** to remove the DC component from Rout **40** (however, in alternate embodiments, a high pass filter may be used instead). The use of DC notch filters may be computationally cheaper than high pass filters and also result in no rippling effect which helps maintain the gain flat through pass band of the filter. In an alternate embodiment, a single shared DC notch filter may be used to perform the functions of DC notch filter **45** and DC notch filter **49**.

Note that adder **36** receives Rin **43** and training signal **41** and provides the sum of the two signals as output Rout **40**; however, if training signal **41** is zero, output Rout **40** is simply the same is input Rin **43**. For the discussions immediately following, it will be assumed that training signal **41** is zero and that Rout **40** is equal to Rin **43**. Also, note that non-adaptive filter **31** and non-adaptive filter **35** are optional and will be discussed further below. For discussions immediately following, it will be assumed that Sin **38** and Sin **39** are equal and error signal **47** is a gain adjusted version of error signal **46**, without the effects of non-adaptive filter **35**.

Sin **39**, therefore, is the send signal which includes any near end talker signal (Sgen) that is transmitted by transmitter **12** and any reflected echo introduced from Rout **40** by hybrid **16**. Therefore, Sin **39** can be expressed as “Sgen+echo”. Adaptive filter **28** provides an estimation of the reflected echo, echo estimation signal **48**, to adder **34**, which outputs error signal **46**. Therefore, error signal **46** can be expressed as “Sin **39**−estimated echo **48**” or, substituting the above expression for Sin **39**, as “Sgen+echo−estimated echo”. When the estimated echo is accurate (i.e. equal or substantially equal to the actual echo), then error signal **46** will include only Sgen without any substantial echo. This is the ideal case. However, if the estimated echo is not accurate, error signal **46** will include both Sgen and a residual echo component. In this case, error signal **46** can be expressed as “Sgen+residual echo” where residual echo is “echo−estimated echo”. When Sgen is absent (that is, when the near end is silent, meaning no signal is being transmitted from transmitter **12**), error signal **46** represents only the residual echo. In this case, error signal **46** may be used to perform an adaptive process to minimize the residual echo, as will be discussed in more detail below. However, if Sgen is present, error signal **46** cannot be used to perform the adaptive process because adaptive filter **28** uses the error to adapt, and with the presence of Sgen, error signal **46** is no longer just the error. Therefore, the detection of Sgen is necessary to determine whether the adaptive process may be performed. Near-End Signal Detector **26**, coupled to receive Sin **38** (which in this example is equal to Sin **39**) and Rin **44**, uses error signal **46** and control signals from monitor and control unit **30** to detect the presence of Sgen (i.e. to detect the presence of a near end talker at transmitter **12**.)

In adaptive filter unit **28**, the echo estimation signal **48**, y(k), is calculated by y(k)=X^{T}(k)·H(k), where X(k)=[x(k), x(k−1), . . . , x(k−N+1)]^{T }is the input signal vector extending over the duration of the FIR filter span; x(n)=Rin **44**. H(k) is a filter coefficient vector for the k-th iteration where H(k)=[h_{0}(k), h_{1}(k), . . . , h_{N−1}(k)]^{T}. The actual update of the filter coefficients is governed by a general LMS-type algorithm: H(k+1)=H(k)+step_size·error(k)·X(k), where error(k) corresponds to error signal **46**; step_size controls the adaptation rate; and H(k+1) is a new filter coefficient vector.

Any residual echo in error signal **46** may further be reduced or removed by nonlinear processor **32**. Nonlinear processor **32** receives error signal **47** (which in this embodiment is a gain adjusted version of error signal **46**) and control signals from monitor and control unit **30** to produce Sout **42**, which, ideally, includes no echo. In addition to reducing or removing the residual echo, nonlinear processor **32** also attempts to preserve or match the background noise of the near end talker signal (Sgen). Matching the background noise allows for improved communication quality by maintaining continuity of the true background noise. Without this continuity, the far end listener may hear only silence from the near end talker when the far end talks. Alternatively, a synthesized background noise may be provided when the far end talks; however, this may result in disruptive switching between true background noise (when the near end talks) and synthesized background noise (when the far end talks). Therefore, matching background noise helps minimize this disruptive switching.

Monitor and control unit **30** includes a filter coefficient monitor (such as filter coefficient monitor **102** which will be discussed further in reference to **28** does not attempt to adapt to invalid hybrids. Monitor and control unit **30** also includes a gain monitor to control gain control **33** within optional adaptive filter **35**. One purpose of gain control **33** is to maintain the stability of communication system **10**. Monitor and control unit **30** also includes a pure delay determinator and a sparse window locator (both of which will be described in more detail with reference to **28**. Monitor and control unit **30** also includes a tone indicator and a tone detector (to be described in more detail with reference to **10**. These signaling tones may include, for example, a 2100 Hz tone with a phase reversal for disabling the echo canceller when data is to be sent following the signaling tone. Therefore, the echo canceller may be disabled as necessary. On the other hand, if adaptive filter **28** is exposed to a tone (such as, for example, a single or multiple frequency sinusoidal) transmitted by either transmitter **12** or transmitter **14**, instability of communication system **10** may result. Therefore, detection of a tone may be used to prevent adaptive filter from diverging and causing instability.

In the embodiments described above, echo canceller **20** did not include non-adaptive filters **31** and **35**. However, in an alternate embodiment, non-adaptive filter **31**, coupled between DC notch filter **45** and adder **34**, can be used to reduce the length of adaptive filter **28** (as will be discussed further in reference to **31** receives Sin **38** and control signals from monitor and control unit **30** to produce Sin **39**. Also, in one embodiment having non-adaptive filter **31**, echo canceller may also include a non-adaptive filter **35** coupled between adder **34** and nonlinear processor **32**. Non-adaptive filter **35** may include gain control **33** or may be a separate unit. In this embodiment, non-adaptive filter **35** compensates the effects of non-adaptive filter **31**, so that the near-end signal Sgen is not distorted. Non-adaptive filter **35** receives error signal **46**, control signals from monitor and control unit **30**, and provides error signal **47** to nonlinear processor **32**. (Non-adaptive filters **31** and **35** will be discussed further below in reference to

Monitor and control unit **30** also provides training signal **41** to adder **36** in order to inject a signal into Rin **43** to produce Rout **40**. The injection of training signal **41** may be used to estimate the pure delay of the hybrid echo path (the path from Rout **40**, through hybrid **16**, and back to Sin **37**). The pure delay refers to the minimum time delay from Rout **40** to Sin **37**. The injection of training signal **41** may be used to estimate the pure delay when the far end signal is absent at the beginning of the communication (such as at the start of a phone conversation). Note that training signal **41** is optional. Monitor and control unit **30** may also receive control **17** to enable or disable all or a portion of the functional modules.

**200** that illustrates operation of echo canceller **20** in accordance with one embodiment of the present invention. Flow **200** is a broad overview of the functionality provided by an echo canceller such as echo canceller **20** of **200** will be provided in more detail below in reference to **10**-**38**. Flow **200** begins at start **202** and flow proceeds to block **204** where DC notch filtering is performed on both Rin and Sin. Note that if adder **36** is present or training signal **41** is present, then DC notch filtering is performed on the output of adder **36** (Rout **40**) rather than Rin **43**. DC notch filter **45**, as mentioned above, removes the DC component from Sin **37** and produces Sin **38**. Similarly, DC notch filter **49** removes the DC component from Rin **43** (or Rout **40**, depending on training signal **41**) and produces Rin **44**. Flow **200** then proceeds to block **206** where long-term power of Rin **44** and short-term power of Sin **38** are estimated. Note that long-term power and short-term power are relative terms. That is, long-term power refers to the power measured over a longer period of time as compared to short-term power. These powers may be calculated by near-end signal detector **26** of echo canceller **20**.

The powers calculated in block **206** are then used to determine a near end talker signal detection (NESD) threshold. This NESD threshold will then be used to determine the existence of a near end talker signal (i.e. Sgen). This determination may also be performed by near-end signal detector **26** of echo canceller **20**. Flow **200** then proceeds to block **210** where adaptive filter **28** is monitored and controlled. Block **210** includes blocks **209**, **211**, and **213**. Note that the functions within monitor and control adaptive filter **210** are optional. That is, any combination of blocks **209**, **211**, and **213** may be performed, or none may be performed. In block **209**, tone indication processing is performed. This tone indication processing may be performed by monitor and control unit **30**, as was described above in reference to **200** then proceeds to block **211** where delay (in one embodiment, pure delay) is detected, and a filtering window with proper size (sparse window) is positioned. That is, monitor and control unit **30** may detect the delay and position the sparse window such that the length (i.e. number of taps) for adaptive filter **28** is reduced.

Another way of shortening adaptive filter length is accomplished by block **213**. One embodiment is to use a combination of non-adaptive filter **31** and **33** in conjunction with adaptive filter **28**, but with a much shorter filter length. Details will be provided in

After monitoring and controlling adaptive filter **210**, flow **200** proceeds to block **212** where an adaptive filter is used to generate an echo estimation signal. For example, this may correspond to adaptive filter **28** generating echo estimation signal **48**, as was introduced above in reference to **200** then proceeds to block **214** where the error signal and the short-term power of the error signal are estimated. That is, block **214** may correspond to adder **34** of **46** by subtracting echo estimation signal **48** from Sin **39**. Monitor and control unit **30** may then be used to estimate the short-term power of error signal **46**.

Afterwards, flow proceeds to block **216** where the NESD threshold is used to detect a near-end talker signal. That is, in block **216**, it is detected whether Sgen exists (whether a signal is being transmitted from transmitter **12** of **26** of **218** where the gain of gain control **33** is monitored and selectively adjusted to maintain stability of adaptive filter **28** and of communication system **10** (the details of which will be described in more detail below). Flow **200** then proceeds to decision diamond **220** where it is determined whether the filter coefficients need to be updated. For example, as discussed above, if Sgen exists, error signal **46** includes both a near-end talker signal (Sgen) and a residual echo component. In this case, adaptive filter **28** should not be updated because error signal **46** is not representative of just the residual echo. Flow then proceeds to decision diamond **224**. However, if it is determined that Sgen does not exist (i.e. that the near-end talker is silent), then the adaptive filter **28** can be updated, and flow proceeds to block **222** where the filter coefficients of adaptive filter **28** are updated prior to continuing to decision diamond **224**.

At decision diamond **224**, it is determined whether any background processing is necessary. In one embodiment, background processing is performed periodically during operation of echo canceller **20**. In alternate embodiments, it can be done at different times, such as in response to various adaptive filter processing states. If background processing is not to be performed, flow proceeds to step **230** where nonlinear processing is performed. However, if background processing is to be performed, flow proceeds to block **226** where the filter coefficients are backed up. That is, the filter coefficients of adaptive filter **28** may be stored (such as in a storage unit which may be located either within echo canceller **20** or external to echo canceller **20**). Flow then proceeds to block **228** where the filter coefficients are monitored to determine whether or not a hybrid exists for echo canceller stability control.

After background processing, if any, flow proceeds to nonlinear processing **230** where any remaining residual echo is reduced or removed and where background noise is inserted, if necessary. If there are more samples being received via Rin **43** and Sin **37** (at decision diamond **232**), processing continues with the next sample back at block **204**, else, the flow is complete at end **234**. Note that in telephony applications, the sampling rate for signals is generally 8 kHz since the signals usually include speech. Therefore, in one embodiment, the sampling rate is 8 kHz, where a sample of Rin **43** and Sin **37** is received every 0.125 ms. However, in alternate embodiments, different sampling rates may be used. For example, a higher sampling rate is generally required for music applications. Furthermore, in digital applications, the sampling rate may depend on the transmission rate of the digital information.

Note that the steps in **200** may be optional, while other embodiments may use additional or different steps to perform any desired operations. Therefore, one of ordinary skill should appreciate that many variations are possible and that flow **200** is only one example of operation of an echo canceller. Similarly, echo canceller **20** also illustrates only one possible embodiment. Alternative embodiments may use more or less blocks or units to perform all, less then all, or even different functions than those illustrated in **20** of

**26**. Operation of near-end signal detector **26** will be described with reference **26** and the flows of **28** in order to minimize the average power of the residual echo) is stopped, as discussed above, to prevent the adaptation from diverging since the existence of a near-end talker signal indicates that error signal **46** is not solely the error due to echo. Note that the adaptation process is stopped when a near-end signal is detected, regardless of whether the near-end signal is during a single-talk situation (i.e. only a near-end talker is present) or a double-talk situation (when both a near-end talker and a far-end talker is present). In addition to stopping the adaptation process, filter coefficients may need to be restored from backed up filter coefficients. Furthermore, when both near-end and far-end signals are absent, the adaptation process is also halted to prevent echo canceller **20** from adapting on channel noise or on low error signals, thus minimizing computation. Therefore, echo canceller **20** operates to adapt when necessary, such as when the far-end signal is relatively strong, and the near-end signal is absent. In this situation, adaptive filter **28** can be adapted to correctly estimate the echo as echo estimation signal **48**. Also, as will be discussed below, the threshold for the near-end talker signal detection is “gear-shifted” (i.e. adjusted), depending upon the state of the adaptive filter process.

The embodiments discussed in **28**. The process may be governed by a state machine, as illustrated in **28** from diverging.

**26**. Signal level estimators track the levels of the near-end signal (Sgen), far-end signal (Rin), and the send path input signal (Sin). Therefore, near-end signal level estimator **50** receives error signal **46**, far-end signal level estimator **52** receives Rin **44**, and Sin signal level estimator **54** receives Sin **38**. The signal level estimations are then used to control near-end signal detection (NESD) threshold selector **58** and near-end signal detector **60**. Background processor **56** monitors the processing status of adaptive filter **28** and controls NESD threshold selector **58** and near-end signal detector **60**. Note that in general, each signal level estimator may apply a low-pass filter on the signal to be measured, and the estimation can be done in either power or magnitude. Also, the following descriptions of FIGS. **3** and **10**-**13** assume that signals are sampled at a rate of 8 kHz (which is a common rate for normal speech applications, as discussed above).

One embodiment of Sin signal level estimator **54** obtains the power of Sin (P_{Sin}) using the following equation:

*P* _{Sin}(*n*)=[(*N−*1)*P* _{Sin}(*n−*1)+(*S*in(*n*))^{2} *]/N* Equation 1:

In the above equation, Sin(n) is the send path input to echo canceller **20** at time n, P_{Sin }(n) is the estimated send path input signal power at time n, and N is a smoothening factor, which, in one embodiment, is assumed to be 32. In alternate embodiments, a range of N values may be used. In general, N should be chosen to be large enough so that the power estimation on Sin is not too sensitive to rapid variations of Sin. On the other hand, N cannot be so large such that the power estimation of Sin is sensitive enough to track the changes of speech signal level, and the delay for the power estimation is minimum. Alternatively, the power can be estimated using a moving average method with window size of 2*N−1 samples. It can be shown that this approach provides equivalent bandwidth to the power estimator as per Equation 1.

Near-end signal level estimator **50** receives error signal **46** and obtains the near-end signal power at time n. As discussed above, though, there is no direct access to the near-end signal (Sgen) for echo canceller **20**. That is, Sin **38** is a mixture of Sgen and the reflected echo from Rin **44**. Therefore, one embodiment of near-end signal level estimator **50** uses the difference between Sin **39** (which is a filtered version of Sin **38**, assuming a filter is present between DC notch filter **45** and adder **34** in **48**. Therefore, error signal **46** is provided to near-end signal level estimator **50**. Error signal **46** is the closest estimation of Sgen available to echo canceller **20**, but the accuracy of this is estimation is a function of the convergence state of adaptive filter **28**. Ideally, when the adaptive filter is fully converged, the estimation of the echo (echo estimation signal **48**) is accurate. In practice, as was described above, echo estimation signal **48** is generally not equal to the reflected echo from Rin **44**, and therefore, error signal **46** is not simply Sgen, but instead is Sgen+residual echo. As the adaptive process continues over a certain window of time, the error introduced by the residual echo is minimized. Therefore, one embodiment of near-end signal level estimator **50** uses the following equation:

*P* _{error}(*n*)=[(*N−*1)*P* _{error}(*n−*1)+(error signal 46)^{2} *]/N* Equation 2:

In the above equation, error signal **46** is the difference between Sin **39** and echo estimation signal **48** at the output of adder **34**, P_{error}(n) is the estimated near-end signal power at time n, and N is a smoothening factor of the estimator (which is 32 in the current embodiment).

One embodiment of far-end signal level estimator **52** obtains a short-term power of Rin and uses this to calculate an average power of Rin over some of the past short-term power estimations of Rin, which covers the range of the echo path. For example, one embodiment determines short-term power of Rin using the following equation:

In the above equation, Rin(kN−i) is the receive path input to echo canceller **20** at time kN−i, P_{Rin}(kN) is the estimated far-end signal power at time kN (note that P_{Rin}(kN) is estimated every N samples, instead of every sample, to reduce computation cost). N is the window size (which is 32 in one embodiment). Therefore, equation 4 calculates the power of Rin within the current window (of size N) every N samples where k keeps track of the windows. That is, the first window (for k=1) may be defined by samples **1**-**32**, the next window (for k=2) may be defined by samples **33**-**64**, etc. The average power of the far-end signal can then be obtained using the following equation:

In the above equation, the “P_{Rin}((k−i)N)” are the past M snapshots of the far-end signal power estimations at time (k−i)N, where i=M−1, M−2, . . . , 0. AVG P_{Rin}(kN) is the average of the far-end signal level estimation at time kN, and M is the window size for the average where M=16 for an echo canceller designed to cover echo path delay up to 64 ms (i.e. M*N=16*32=512 samples). For example, if the current window is 16^{th }window (i.e. k=16), then the value of AVG P_{Rin}(kN) takes the average of the P_{Rin}(kN) values that were calculated for each of the 16 (i.e. M) previous windows, i.e. the average of: P_{Rin}(16*N), P_{Rin}(15*N), . . . , P_{Rin}(2*N), P_{Rin}(N). If there is not enough previous data (i.e. less than M windows have been processed), then only those available values can be used to determine the average and a zero may be utilized for values that have not yet been calculated. For example, for the 3^{rd }window (k=3), only two previous P_{Rin}(kN) values are available, and therefore, AVG P_{Rin}(kN) would be an average of 3 values and not M values. Note also that AVG P_{Rin}(kN) is calculated every N samples, as will be seen in reference to _{Rin}(kN), the value of k can be incremented to indicate the start of a next window of N samples.

In an alternate embodiment, far-end signal level estimator **52** can estimate the average power of Rin using either equation 1 or 2 above with N=256. As with equation 4, the measurements AVG P_{Rin}(kN) should also be taken every 32 samples. Note also that all the above level estimations can be done using magnitude rather than power. Also, note that equations 1 and 2 above process data at a nominal rate while equations 3 and 4 perform sub-rate calculations on sequential N-size windows (thus performing calculations once every N samples). Alternate embodiments may structure the above equations in any manner and are not restricted to those equations given above.

The level estimations for the near-end (Sgen), far-end (Rin), and send path input (Sin) signals are used in the control of the near-end talker signal detection. **206** and **208** of **250**, the short-term power of Rin, P_{Rin}, is determined (e.g. see equation 3 above) where the current time, n, corresponds to the current sample being analyzed. In block **252**, the sample counter is incremented, thus providing a new value of n. The sample counter is therefore incremented during each pass through flow **200** of **206** and **208**).

Flow proceeds to decision diamond **254**, where it is determined whether the sample counter has reached the window size, N. If not, flow proceeds to block **270** (and note that the NESD_Threshold is not updated). In block **270**, the power of Sin (P_{Sin}) is calculated (e.g. see equation 1 above). If the sample counter has reached N, then flow proceeds to block **256** where the counter is reset such that the path of blocks **256** through **268** is only taken every N samples (which in one embodiment occurs every 32 samples with N being 32). After the sample counter is reset (reset to zero, in one embodiment), flow proceeds to block **258** where the average power of Rin, AVG P_{Rin}, is calculated (e.g. see equation 4 above). Afterwards, flow proceeds to decision diamond **260** where it is determined whether BACKUP_STATE is 0 or 1 (note that BACKUP_STATE will be described in more detail in reference to **262** where K**1** is used to adjust NESD_Threshold as shown below in equation 5:

*NESD*_Threshold=*K*1*AVG *P* _{Rin} Equation 5:

However, if BACKUP_STATE is not 0 or 1, then flow proceeds to block **264** where K**2** is used to adjust NESD_Threshold as shown below in equation 6:

*NESD*_Threshold=*K*2*AVG *P* _{Rin} Equation 6:

Therefore, NESD_Threshold is reevaluated once every N samples, and depending upon the state of the adaptive filters (i.e. BACKUP_STATE, corresponding to a state machine which will described further in reference to **1***AVG P_{Rin }or K**2***AVG P_{Rin}. K**1** and K**2** are NESD threshold scaling factors. During the initial phase of the adaptation process of adaptive filter **28** (i.e. when BACKUP_STATE is 0 or 1), NESD_Threshold can be relatively large, providing more opportunity for adaptive filter **28** to adapt. On the other hand, when adaptive filter **28** has passed the initial adaptation phase (i.e. when BACKUP_STATE is 3 or 4), NESD_Threshold may be reduced to prevent the adaptation process from diverging. In one embodiment, K**1** is set to a value within a range of 1 to 2 while K**2** is set to a value within a range of 0.25 to 1. For example, in one embodiment, K**1** is 1 and K**2** is 0.5, depending upon hybrid conditions. Furthermore, these values of K**1** and K**2** can be set either statically or dynamically during the adaptive process. Alternate values outside the ranges given above may be used, and other methods, other than the use of a state machine having 4 states (e.g. BACKUP_STATES 0-3) may be used to determine when the adaptation process is still in its initial phase.

After adjusting NESD_Threshold in block **262** or **264**, flow proceeds to decision diamond **266** where it is determined whether NESD_Threshold is less than A. If not, flow proceeds to block **270** where the P_{sin }is calculated. If so, NESD_Threshold is set to A in block **268**. That is, NESD_Threshold is limited at a minimum level of A (where, in one embodiment, A may correspond to a value in a range of −40 to −45 dBm0) in the case where AVG P_{Rin }is too small. Flow then proceeds to block **270**.

After the flow of **28** is performed (see block **210** of **209**, **210**, and **211** will be described in more detail in reference to **48** in block **212**, and flow proceeds to blocks **214** and **216** which are illustrated in further detail in reference to **11** illustrates a portion of blocks **214** and **216** of

In block **276**, the power of error signal **46** (P_{error}) is estimated (see equation 2 above). Flow then proceeds to decision diamond **278** where it is determined whether the smaller of P_{error }and P_{Sin }is greater than NESD_Threshold (i.e. whether MIN(P_{error}, P_{Sin})>NESD_Threshold). If so, flow proceeds to decision diamond **280** where it is determined whether NESD_Hangover timer has counted down to zero. If it has, then a near-end signal has been detected. That is, a near-end signal is detected only when MIN(P_{error}, P_{Sin})>NESD_Threshold and no near-end signal has been detected during a certain time window in the past (corresponding to the NESD_Hangover timer). If at decision diamond **278**, the MIN(P_{error}, P_{Sin}) is not greater than NESD_Threshold, flow proceeds to block **290** where the value of the NESD_Hangover timer is decremented until it reaches zero, thus introducing a pause determined by the NESD_Hangover time. If at decision diamond **280**, the NESD_Hangover timer is not zero, the NESD_Hangover timer is set to a predetermined value in block **286**.

If, a near-end signal (Sgen) has been detected, flow proceeds from decision diamond **280** to decision diamond **282** where it is determined whether the filter coefficients have been updated. If so, it is assumed that the coefficients are mostly likely corrupted due to the presence of a near-end signal. That is, because the signal being used for the coefficient update is no longer the pure residual echo, but a mixture of the residual echo and Sgen, the coefficients are no longer representative of the estimated echo. In this case, flow proceeds to block **284** where the filter coefficients are restored or replaced by a proven “good” set of filter coefficients. The method of backing up and restoring the filter coefficients will be described below in reference to **288** where BACKUP_STATE is updated. If the filter coefficients have not been updated at decision diamond **282**, then the coefficients are not assumed to be corrupted because they have not been adapted using a mixture of residual echo and Sgen. In this case, flow proceeds to block **286** where the NESD_Hangover timer is set to the predetermined value.

The duration of the NESD_Hangover time that is used for the NESD_Hangover timer is chosen to ensure that Sgen is no longer present before starting filter coefficient adaptation as well as to avoid any unnecessary filter coefficient adaptation and restore. For example, in one embodiment, the NESD_Hangover time is 160 samples, or 20 milliseconds. Therefore, the duration of the NESD_Hangover time prevents near-end signal detector **26** from being overly sensitive thus minimizing the switching between the detection of a near-end talker signal and detection of a lack of near-end talker signal. However, if the NESD_Hangover time is set too long, near-end signal detector **26** may not be sensitive enough to accurately detect a near-end talker signal when necessary.

Therefore, under different combinations of the signal levels (i.e. power) of Sgen and Rin, different action regarding the filter coefficients (e.g. the coefficients of adaptive filter **28**) is taken. For example, these actions can be summarized using the following table:

TABLE 1 | |||||

Item | Sgen | Sin | Rin | Action | Description |

1 | low | low | low | no update | no near-end or far-end talker signals |

present | |||||

2 | low | low | high | update coefficients | single far-end talker signal with |

large hybrid attenuation | |||||

3 | low | high | low | n/a | not a valid combination |

4 | low | high | high | update coefficients | single far-end talker signal with |

small hybrid attenuation | |||||

5 | high | low | low | n/a | not a valid combination |

6 | high | low | high | n/a | not a valid combination |

7 | high | high | low | freeze update and restore | single near-end talker signal present |

filter coefficients | |||||

8 | high | high | high | freeze update and restore | double-talk (both near-end and far- |

filter coefficients | end talker signals present) | ||||

Since Sin is a mixture of Sgen and the echo or Rin, several combinations listed in the above table are not valid ones under normal operation mode (meaning that the connection is not broken or that no extra signals are injected into the circuit). These invalid combinations are Item 3 (because Sin cannot be high when both Sgen and Rin are low), Items 5 and 6 (because Sin cannot be low when Sgen is high). Three sets of actions are used for the remaining 5 combinations. Firstly, the condition for the coefficient adaptation process is when Rin is high while Sgen is low (during a single far-end talking period, regardless of whether Sin is high or low, i.e. items 2 and 4, respectively). Under these conditions, the error (error signal **46**) is due mainly to residual echo (because Sgen is low), and the effect of Sgen on adaptation is minimal. Secondly, the conditions for stopping the filter coefficient adaptation processor and for restoring previously determined “good” filter coefficients are when Sgen is high (during a near-end talking period, regardless of single talk or double talk, i.e. items 7 and 8, respectively). Thirdly, no update is necessary when both the near-end and the far-end talkers are silent (item 1).

The methods described above for the detection of Sgen allows for the ability to cancel echoes when the echo return loss is close to or less than 6 dB. Therefore, at close to or less than 6 dB (such as in item 4 of Table 1 above), the methods above use the minimum of Sgen (which, as described above, may be estimated as error signal **46**, assuming that the residual error is small or negligible) and Sin which results in no false detection under these conditions, unlike previous solutions which may be heavily affected by variations in echo return loss at these lower levels (closer to or less than 6 dB) and therefore tend to falsely detect the presence of Sgen when it is actually not present. Furthermore, the methods above enable the adaptation process to continue when the echo return loss is up to 0 dB (no hybrid attenuation), allowing the echoes to be cancelled, unlike in prior solutions where the adaptation process was stopped at levels such as 6 dB.

Also, the methods described above in reference to **46**, assuming that the residual error is small or negligible) and Sin as compared with Rin for the near-end signal detection, the detection threshold (NESD_Threshold) setting is independent of the echo return loss resulting in a near-end signal detection that is faster and more reliable than previously available solutions.

Furthermore, the methods described above in reference to

**224** and block **226** of **28**) backup policy. One embodiment of the backup policy ensures that the good filter coefficients are being backed up periodically, to minimize the number of backups, and to minimize the frequency of backups. **291** where the background **1** counter is incremented. Flow proceeds to decision diamond **293** where it is determined whether the background **1** counter has reached a predetermined counter value, J. If not flow proceeds to point H (after block **228** in **298** where background **1** counter is reset (to zero), and then to decision diamond **295** where it is determined whether the filter coefficients of adaptive filter **28** have been updated. If not, flow proceeds to point H. If so, flow proceeds to block **292** where the background **2** counter is incremented.

Flow then proceeds to decision diamond **294** where it is determined whether the background **2** counter has reached a predetermined counter value, L. If not, flow proceeds to point H. If so, flow proceeds to block **296** where background processing is performed. That is, background processing, in this embodiment, is performed at most every J*L samples, and these values, J and L, can be set to any value which helps to determine the frequency of background processing. For example, in one embodiment, J is 160 samples and L is 10, where background processing is performed at most every J*L, or 1600, samples. That is, if, after J samples, the filter coefficients of adaptive filter **28** have not been updated, then flow proceeds to point H, and the background **2** counter is not incremented. Therefore, the background **2** counter is incremented and compared to L only if the coefficients have been updated during the current window of J samples. In block **296**, the background **2** counter is reset (in this embodiment, reset to zero). Flow continues from block **296** to decision diamond **300**.

In decision diamond **300**, it is determined whether the current BACKUP_STATE (which will be described in more detail in reference to **304** and flow proceeds to block **308**. If the BACKUP_STATE is not 0 or 1, flow proceeds to decision diamond **302** where it is determined whether BACKUP_STATE is 2. If not, flow proceeds to block **306** (indicated that BACKUP_STATE is 3) where BACKUP_STATE is set to 2 and flow proceeds to block **310**. If BACKUP_STATE is 2 at decision diamond **302**, flow proceeds to block **308** where the Candidate backup coefficients are copied to the Good backup coefficients. (Note that Candidate and Good backup coefficients will be described below in reference to **310** where the current filter coefficients are copied to the Candidate backup coefficients. That is, in block **308**, the Candidate backup coefficients become the Good backup coefficients, and the current filter coefficients become the Candidate backup coefficients, where the current backup coefficients, Candidate backup coefficients, and Good backup coefficients can all be stored in a storage unit or separate storage units either in echo canceller **20** or in storage location outside of echo canceller **20**. Afterwards, flow proceeds to block **228** of

One embodiment of the present invention uses two coefficient backups marked as Candidate backup coefficients and Good backup coefficients, and has a combination of 4 different BACKUP_STATES (0 to 3). **28**.

The state machine of **216** and **226** of

In one embodiment, the state machine starts with STATE 0 upon reset or initialization. The state machine transitions to STATE 1 if no near-end signal (Sgen) is detected in the last L entries to the background processing. Therefore, the minimum time window for the first backup of the filter coefficients is J*L samples (where L is 10 in the current embodiment, and J*K is therefore 1600 samples or 200 ms, assuming a sampling rate of 8 kHz). For this state transition, near-end signal has not been detected, and a first backup is performed by copying the current filter coefficients to the Candidate backup coefficients. Upon the detection of a near-end signal (Sgen), the state machine transitions back to STATE 0 because the stored Candidate backup coefficients may be corrupted due to the delay in the detection of the near-end signal, Sgen. The state machine remains in STATE 0 until no near-end signal is detected in the last L entries to the background processing at which point, the state machine again transitions to STATE 1 as was described above.

In STATE 1, if no near-end signal is detected in another L entries of the background processing, the state machine transitions to STATE 2 where a second backup is performed by copying the Candidate backup coefficients to the Good backup coefficients and copying the current filter coefficients to the Candidate backup coefficients. In this state, both the Candidate and Good backup coefficients are available and the state machine will remain in this state if no near-end signal is detected. Note that in one embodiment, both Candidate and Good backup coefficients are renewed in sequential backups during the second backup, even though the state is not changed. Also, in one embodiment the two copies performed upon transition to STATE 2 from STATE 1 are performed with a single copy by first marking the Candidate backup coefficients as the Good backup coefficients (through the use of a pointer, for example), and then copying the current filter coefficients to the Candidate backup coefficients (which used to be marked as the Good backup coefficients).

In STATE 2, when a near-end signal is detected, the state machine transitions to STATE 3 where the Candidate backup coefficients are again considered corrupted but the Good backup coefficients are still considered good because these Good backup coefficients have been proven to be good after at least a J*L time window. The state machine remains in STATE 3 so long as the near-end signal persists, or will go back to STATE 2 if the near-end signal is no longer present.

Note that in alternate embodiments, each entry to the background processing (L) can occur on each sample rather than every J samples. Also, the state machine of

**30** of **33** of **10** and of adaptive filter **28**. For example, system **10** is considered unstable if it produces sustained artifacts due to a set of filter coefficients (of adaptive filter **28**) that are very different from the impulse response of hybrid **16**. As mentioned above with respect to **28** attempt to “imitate” the impulse response of hybrid **16** and subtract it out from the outgoing signal in an attempt to cancel out the reflected echo. However, if the coefficients of adaptive filter **28** vary too much from the impulse response, artifacts such as voice or data signal distortions or even system howling may occur. The instability of system **10** can occur under the following two conditions: (1) echo cancellers **20** and **22** being in a closed-loop system and stimulated by certain type of signals thus resulting in a gain of greater than 1 for system **10** and (2) echo canceller **20** being in an open-loop system.

**20**, which may be performed by gain monitor **100** which is coupled to adaptive filter **28** and gain control **33**. The dynamic gain-control method of **20** and **22** in a closed-loop system. For example, if error signal **46** is greater than Sin **38** (which in theory should not occur, but in practice can occur), the gain of echo canceller **20** is greater than or equal to one. If the same happens in echo canceller **22** (resulting in a gain of echo canceller **22** also being greater than or equal to one), then the entire loop gain of the closed-loop system with echo cancellers **20** and **22** may be greater than one which can produce an artifact known as howling. Therefore, the method of **46** when the ratio of the power of error signal **46** (P_{error}) versus the power of Sin **38** (P_{Sin}) within a certain time window (see equations 1 and 2 above) is larger than an adaptive threshold. In addition, the method of **28** when P_{error }is many times larger than P_{Sin}. The method of **10**. Furthermore, the method of **28** upon a sudden change in hybrid characteristics.

Therefore, **218** of **216** of **218** where the gain of echo canceller **20** is monitored and selectively adjusted. Therefore, flow begins with decision diamond **322** where it is determined whether the ratio of P_{error }to P_{Sin }(P_{error}/P_{Sin}) is greater than a reset threshold. If so, flow proceeds to block **330** where the filter coefficients of adaptive filter **28** are reset (i.e. set to zero, in one embodiment). Alternatively, the coefficients can be reset to any value. Therefore, the reset threshold may be used to determine whether P_{error }is too much greater than P_{Sin}, thus requiring the reset of adaptive filter **28** to prevent instability. The reset threshold can therefore be any value, and in one embodiment is set to 8.

If P_{error}/P_{Sin }is not greater than the reset threshold, flow continues to decision diamond **324** where it is determined whether P_{error}/P_{Sin }is greater than a gain threshold. The gain threshold is generally less than the reset threshold and in one embodiment, is set to 1. This gain threshold is a threshold for starting activation of gain attenuation. If P_{error}/P_{Sin }is greater than the gain threshold, flow proceeds to block **328** where the gain is adjusted using alpha, as shown in equation 7 below:

gain=alpha*gain Equation 7:

Alpha is generally less than 1 such that error signal **46** is attenuated. Therefore, in one embodiment, alpha is 0.9996. Flow proceeds to decision diamond **328** where it is determined whether the gain is less than a gain limit. If so, flow proceeds to block **334** where the gain is set to a gain limit. This ensures that the gain never falls below a predetermined level, which in one embodiment, is 0.5. For example, it is generally not desirable to cut off the send path transmission path completely (i.e., gain=0), even under some abnormal situations, such as the hybrid being in an open-loop circuit. Flow then proceeds to block **326**. If, at decision diamond **332** it is determined that the gain is not less than the gain limit, flow proceeds to block **326** where error signal **47** is calculated as shown in equation 8 below:

error signal 47=gain*error signal 46 Equation 8:

If, at decision diamond **324**, it is determined that P_{error}/P_{Sin }is not greater than the gain threshold, flow proceeds to decision diamond **336** where it is determined whether the gain is less than 1. If not, flow proceeds to block **326** where error signal **46** is attenuated; however, if it is less than 1, then flow proceeds to block **338** where the gain is adjusted as shown below in equation 9:

gain=beta*gain Equation 9:

Beta is generally greater than 1 because since the gain was previously attenuated, it needs to be recovered. Therefore, in one embodiment, beta is 1.0004. Flow then proceeds to decision diamond **340** where it is determined whether the gain is greater than 1. If so, flow proceeds to block **326** where error signal **46** is attenuated, and if not, flow proceeds to block **342** where the gain is set to 1. After block **342**, flow proceeds to block **326** where error signal **46** is not attenuated because error signal **47** is simply equal to error signal **46***1 (since the gain was set to 1 in block **342**). Therefore, in summary, if P_{error}/P_{Sin }is greater than or equal to the reset threshold, the filter coefficients of adaptive filter **28** are reset. If P_{error}/P_{Sin }is less than the reset threshold but greater than or equal to the gain threshold, then the error is attenuated by the gain value (e.g. in block **326**). However, if P_{error}/P_{Sin }is also less than the gain threshold, then the error is left unattenuated (i.e. error signal **47**=error signal **46**). Therefore, it can be appreciated how the flow of **10**.

**28**, which may be performed by filter coefficient monitor **102** within monitor and control unit **30** and coupled to adaptive filter **28**. The method of **20** in an open-loop system. The monitoring method detects the formation of a set of filter coefficients of adaptive filter **28** having a relatively uniform distribution. Since an impulse response by hybrid **16** is expected, a uniform distribution of the coefficients of adaptive filter **28** indicates that no hybrid exists, thus indicating the possibility of an open-loop condition. Therefore, upon detecting a uniform distribution of the coefficients of adaptive filter **28**, the filter coefficients are reset, and echo canceller **20** is placed in an alert state for further monitoring. When the filter coefficients are reset repeatedly during a certain time window, it is assumed that echo canceller **20** is in an open-loop condition and echo canceller **20** is bypassed. That is, adaptive filter **28** should only adapt if a true hybrid exists. Furthermore, adaptive filter **28** in an open-loop system with continuous sinusoidal inputs via Rin and non-zero signals as Sin (e.g. sinusoidal tones) may diverge especially fast, thus increasing the need for the detection of an open-loop system.

Therefore, **228** of **226** of **228** where the coefficients of adaptive filter **28** are monitored. Therefore, flow begins with block **344** where the filter coefficients of adaptive filter **28** are divided into B number of bins. (B is selected to be number of the filter coefficients/16.) Flow proceeds to block **346** where the maximum and minimum coefficients power of the B bins is determined. That is, if the filter coefficients are divided into B bins, each bin will have associated with it a power value of the coefficients within that bin (e.g. an average power of the coefficients within that bin), and in block **346**, a maximum power value of the B bins and a minimum power value of the B bins is selected. Flow continues to decision diamond **328** where it is determined whether a ratio of the maximum power value and the minimum power value (i.e. maximum power/minimum power) is less than an alert threshold. If the filter is adapted towards a real hybrid, the ratio of the maximum power over the minimum power should be far greater than 1. On the other hand, if the ratio of the maximum power over the minimum power is close to 1, it is a clear indication that the filter is not adapting to a real hybrid. A ratio is chosen as an alert threshold for signaling the possibility of the absence of a hybrid. The alert threshold is chosen based on statistical analysis of the adaptive filter behaviors under various hybrids. In one embodiment, the alert threshold is chosen to be 8.

After the comparison, flow continues to block **350** where the filter coefficients of adaptive filter **28** are reset to zero (or set to any other predetermined reset value or values). Flow continues to block **352** where the alert state is incremented. (The alert state indicates how many times the filter coefficients have been reset during the current period of time in which the ratio of maximum power to minim power is less than the alert threshold. Note that the current period of time is the same J*L as was discussed above with reference to **310** of **228** of **344** of **352**, flow proceeds to decision diamond **354** where it is determined whether the alert state is equal to a bypass threshold. If not, then echo canceller **20** is not placed in bypass mode and therefore adaptive filter **28** continues to adapt. However, if alert state has reached the bypass threshold in decision diamond **354**, flow proceeds to block **356** where bypass mode is set to 1 indicating that an open-loop condition has been detected (i.e. no hybrid exists) and therefore echo canceller **20** is to be bypassed so as not to adapt to a non-existent hybrid.

If, at decision diamond **348**, it is determined that the ratio of maximum power to minimum power is not less than the alert threshold, flow proceeds to block **358** where the alert state is reset to 0. Flow proceeds to decision diamond **360** where it is determined whether bypass mode is 1 and if so, it is reset to 0 in block **362**. The branch to **358** therefore allows for a reconnection of hybrid **16** where adaptive filter **28** begins to adapt again.

**32** of **32** also attempts to preserve or match the background noise of the near-end talker signal which allows for improved communication quality. In general, nonlinear processor **32** detects if the residual echo is below a certain threshold and replaces it with comfort noise, rather than silence, to avoid a sudden disappearance of the telephone line background noise. Such sudden disappearance of background noise may lead to an impression that the telephone connection has been broken.

One prior art method used today uses a synthesized background noise; however, this may result in disruptive switching between true background noise and the synthesized background noise. For example, one prior art method used today uses white noise as comfort noise. However, white noise is far different from natural background noise and therefore sounds disruptive. An alternate solution available today repeatedly outputs pre-stored background noise signals to match background noise. However, this method requires additional storage space and results in the noticeable repetition of background noise which may also be disruptive to communication.

Therefore, **32** which preserves or matches natural background noise in echo canceller **20** in order to reduce artifacts caused by the nonlinear processing of echo cancellation such as the disruptive artifacts discussed in the previous paragraph. Nonlinear processor **32** utilizes short term level estimator **88** and long term signal level estimator **92** to find a reliable estimation for the level of the true background noise signals, and to adjust its thresholds (NLP_ON and NLP_OFF thresholds, to be discussed below. The short-term estimator produces a rapid level estimation of the background noise signals at the beginning of a call. The long term estimator, on the other hand, is adaptive in nature aiming at reliably tracking the background noise signal level over time. A decision of activating nonlinear processor **32** is made based on the relative levels of the far-end signals, the near-end signals, and the background noise signals. When the background noise signals become noticeable, nonlinear processor **32** preserves the original background noise signals by passing them through echo canceller **20**. When the background noise signals are low and the residual echo becomes audible, nonlinear processor **32** replaces the residual echo with comfort noise signals of a level a couple of dB lower than the estimated background noise signal level. The generated comfort noise signals are also gradually blended into the original background noise signals to minimize the transition audibility. Therefore, nonlinear processor **32** preserves the natural background noise when possible or matches the background noise with minimum audible effects.

The preservation or matching of natural background noise in echo canceller **20** is performed in four basic steps: (1) estimating the levels of the background noise signals, the far-end talker signals, and near-end talker signals; (2) determining the thresholds for nonlinear processor **32**; (3) generating comfort noise if nonlinear processor **32** is needed; and (4) mixing the comfort noise into the background noise if nonlinear processor **32** is needed.

Nonlinear processor **32** of **96** which includes short-term background level estimator **88**, background level estimator controller **90**, long-term background level estimator **92**, and background level adapter **94**. The estimation for the background noise level is done by short-term background level estimator **88** and long-term background level estimator **92**. Short-term background level estimator **88** provides the initial rapid estimation when opening a call, and long-term background level estimator **92** gradually adapts to the level of the background noise signals over time. Note that the adaptation rate of long-term background level estimator **92** to a higher noise level is slower than the adaptation rate to a lower noise level when the background noise level changes. Therefore, estimators **88** and **92** are active when both the levels of the near-end and far-end talker signals are below predetermined thresholds. That is, if a values are available for a long-term background level estimation, only estimator **92** is used. Therefore, short-term background level estimator **88** is generally only used at the beginning (i.e. at the beginning of a call) when long-term background level estimator **92** is not available yet. (The levels of the near-end and far-end talker signals are determined by near-end signal level estimator **70** and far-end signal level estimator **72**, respectively.)

The threshold for turning on nonlinear processor **32** (performed by nonlinear processor ON controller **76**) is different than the threshold for turning it off (performed by nonlinear processor OFF controller **78**). Nonlinear processor ON controller **76** enables (or turns on) nonlinear processor **32** when the near-end talker signals are insignificant and the far-end talker signals are active. Nonlinear processor OFF controller **78** disables (or turns off) nonlinear processor **32** when the near-end talker signals are relatively high, or the background noise signals are very noticeable. The trade-off between eliminating the residual echo and preserving the actual background noise is made as follows. When the background noise signals are relatively high, nonlinear processor **32** is disabled to allow the background noise to pass through echo canceller **20**. In this case, the negligible residual echo is buried by the much noticeable background noise signals, due to a masking effect. When the background noise signals are relatively low, nonlinear processor **32** is enabled because the residual echo is more audible when it is present with rather quiet background noise signals. In both cases, through, the residual echo is small due to good convergence depth achieved by adaptive filter **28**.

When nonlinear processor **32** is enabled, comfort noise is generated (by comfort noise generator **86**) and the noise levels are matched (by noise level matcher **82**) to minimize the audible “noise gating” (i.e. noise switching from one background to another or from one background to silence) for the perceived speech. Several types of comfort noise signals may be chosen to be close to natural background noise signals. In addition, the comfort noise gradually replaces the actual background noise (performed by output signal mixer **84**) to smoothen the transition, and the level of the comfort noise is set to be a couple of dB lower than the estimated background noise level.

**400**, **402**, and **404** of **400**, it is determined whether the level of the near-end talker signals (P_{error}) are below an error power threshold. The error power threshold is defined as a threshold to determine whether the error signal is considered as the background noise signal, or near-end talker signal. In one embodiment, the error threshold is −39 dBm0. This check reduces the likelihood of mixing the near-end talker signals with the background noise signals, because the background energy estimation to be described below cannot include the near-end talker signals. If P_{error }is less than the error threshold, flow proceeds to decision diamond **402** where the second condition is checked. In decision diamond **402**, it is determined whether the level of the far-end talker signals (P_{Rin}) are less than an Rin threshold in order to exclude the residual echo in the background level estimation. The Rin threshold is defined as an Rin signal level significant enough to generate noticeable residual echo before the non-linear processor. In one embodiment, Rin threshold is −27 dBm0. If P_{Rin }is less than Rin threshold, flow proceeds to decision diamond **404** where it is determined whether the first two conditions have been met for a certain time window (i.e. the background hangover time). That is, if background hangover timer=0, then the first two conditions have been met for the time window defined by background hangover time, and flow proceeds to block **408**. The background hangover time is used to ensure that the far- and the near-end talker signals have been absent for a certain time window. In one embodiment, the background hangover time is 160 samples, or 20 ms, assuming a sampling rate of 8 kHz.

If P_{error }is not less than the error threshold at decision diamond **400** or if P_{Rin }is not less than the Rin threshold at decision diamond **402**, flow proceeds to block **406** where the background hangover timer is set to a predetermined value, e.g. the background hangover time discussed in the previous paragraph. Then flow proceeds to point C. (Note that at point C, flow continues to **404**, the background hangover timer is not 0, then the background hangover timer is decremented in block **410** and flow proceeds to point C.

However, when the **3** conditions of decision diamonds **400**, **402**, and **404** are met, flow proceeds to block **408** where the background level (P_{background}) is adapted to a desired one determined in a later step (P_{new} _{ — } _{background}). (Note that P_{new} _{ — } _{background }will be calculated and discussed in reference to block **426** in **408**, P_{new} _{ — } _{background }may have any appropriate initial value, such as an initial value representative of a comfort noise level.) The adaptation is done for every sample to smooth the transition from one signal level to another in the comfort noise level matching. Therefore, the adaptation is performed as shown in equation 10 below.

*P* _{background}(*n*)=[(*R−*1)*P* _{background}(*n−*1)+*P* _{new} _{ — } _{background} *]/R* Equation 10:

In equation 10, P_{background}(n) is the estimated background power level at time n; P_{new} _{ — } _{background }is the new background power level to be adapted (and is determined in the fourth step); and R is a factor controlling the adaptation rate, which is set to either FAST_RATE or SLOW_RATE. (Note that R may be set in block **428** of **480**, **472**, or **476** of ^{9 }and for SLOW_RATE is set as 2^{11}.)

After block **408**, the estimation of the power level of the background noise signal begins, which includes 3 major steps. The first step in estimating the power level of the background noise signals is to calculate the background power level within a window. Therefore, flow proceeds to block **412** where the power of a windowed background (P_{window} _{ — } _{background}) is calculated as shown below in equation 11.

In equation 11, P_{window} _{ — } _{background }is the windowed background power level estimation, error signal **46** is the difference between Sin **39** and echo estimation signal **48** at the output of adder **34** of **414** where the background sample counter is incremented.

The second step includes finding the minimum P_{window} _{ — } _{background }over a certain number of time windows, w_count. (In one embodiment, w_count is 128 samples; however, in alternate embodiments, w_count can be any value depending on the number of time windows desired for calculating the minimum P_{window} _{ — } _{background}.) Therefore, the calculation of block **418** (shown in equation 12 below) is performed once every w_size samples. For performing the second step, flow proceeds to decision diamond **416** where it is determined whether the background sample counter is w_size. If not, flow proceeds to point C (in **418** where the minimum power of windowed background is determined as shown in equation 12 below.

*P* _{min} _{ — } _{window} _{ — } _{background}=MIN(*P* _{old} _{ — } _{min} _{ — } _{window} _{ — } _{background} *, P* _{window} _{ — } _{background}) Equation 12:

Therefore, P_{min} _{ — } _{window} _{ — } _{background }is determined by selecting the minimum between the old minimum power (the minimum power determined during the previous iteration through block **418**) and P_{window} _{ — } _{background }determined in block **412**. Flow then proceeds to block **420** where P_{window} _{ — } _{background }is reset to zero. Flow proceeds to block **422** where the background sample counter is reset to 0 and the window counter is incremented. Flow then proceeds to point A which continues with **424**).

The third step in the adaptive background level estimation is to determine P_{new} _{ — } _{background }for the background level adaptation discussed in reference to block **408** and to determine the adaptation rate used in block **408**. There are two different approaches depending upon whether it is the first time to determine P_{new} _{ — } _{background}. Therefore, in decision diamond **424** it is determined whether this is the initial estimation (indicating no long-term data is available, such as at the beginning of a call). If so, flow proceeds to block **426** where P_{new} _{ — } _{background }is set to the P_{window} _{ — } _{background }calculated in the first step. Flow then proceeds to block **428** where the adaptation rate R is set to FAST_RATE. However, if at decision diamond **424** it is determined that this is not the initial estimation (indicating that P_{new} _{ — } _{background }is already available because long term data, e.g. N previous samples, is available), flow proceeds to decision diamond **430**. Note that if it is not the initial estimation, the process of determining P_{new} _{ — } _{background }is done once every w_count windows. Therefore, at decision diamond **430**, it is determined whether the window counter has reached w_count. If not, flow proceeds to point C (in **432** where is P_{new} _{ — } _{background }calculated. Flow then proceeds to block **434** where the adaptation rate R is determined. (The details of the determinations of P_{new} _{ — } _{background }and R will be described further in reference to **436** where the window counter is reset to 0 and then to block **438** where P_{min} _{ — } _{window} _{ — } _{background }is reset to 0. Flow then proceeds to point C.

_{new} _{ — } _{background }and R when P_{new} _{ — } _{background }is available. The method of _{new} _{ — } _{background }having a large jump from a lower level to a higher level but places no such constrain when the change is from a higher level to a lower level since this change is faster. Therefore, in one embodiment P_{new} _{ — } _{background }is capped to be no more than two times P_{background}. The method of

In **432** and **434** of **466** where it is determined whether P_{min} _{ — } _{window} _{ — } _{background }is greater than a constant times P_{background}, i.e. whether “P_{min} _{ — } _{window} _{ — } _{background}>constant*P_{background},” where, in one embodiment, the constant is 0.5. If so, flow proceeds to block **478** where P_{new} _{ — } _{background }is set to “(constant*P_{min} _{ — } _{window} _{ — } _{background})+comfort noise level”. In one embodiment, the constant in block **478** is 2 (where this 2 corresponds to the 0.5 of the previous sentence). Flow proceeds to block **480** where the adaptation rate is set to SLOW_RATE. Flow then proceeds to block **436** of

If at decision diamond **466**, P_{min} _{ — } _{window} _{ — } _{background }is not greater than “constant*P_{background},” then flow proceeds to decision diamond **468** where it is determined whether P_{min} _{ — } _{window} _{ — } _{background }is greater than P_{background}. If so, flow proceeds to block **474** where P_{new} _{ — } _{background }is set to P_{min} _{ — } _{window} _{ — } _{background}. Flow then proceeds to block **476** where the adaptation rate R is set to SLOW_RATE. Flow then proceeds to block **436** of **468** it is determined that P_{min} _{ — } _{window} _{ — } _{background }is not greater than P_{background}, then flow proceeds to block **470** where P_{new} _{ — } _{background }is set to “P_{min} _{ — } _{window} _{ — } _{background}+comfort noise level”. Flow then proceeds to block **472** where the adaptation rate R is set to FAST_RATE and then to block **436** of

Therefore, note that comfort noise level (CNL) is added (in blocks **478** and **470**) in order to prevent P_{new} _{ — } _{background }from being silent, when P_{background }or P_{min} _{ — } _{window} _{ — } _{background }happens to be 0. For example, in one embodiment, CNL is set to −66 dBm0. Alternatively, CNL can be in a range of −60 to −72 dBm0. Also, although the flow of

**230** of **406**, block **410**, or decision diamond **416** of **438** in **440** where it is determined whether P_{error }is greater than the nonlinear processor off (NLP_OFF) threshold. If so, flow proceeds to block **452** where NLP_OFF is set (indicating that nonlinear processor **32** is turned off) and then to block **454** where the noise ramping factor is reset to a predetermined value. The noise ramping factor is used to smoothen the signal level transition from low to high. (After block **454**, flow proceeds to block **232** of **440**, it is determined that P_{error }is not greater than the NLP_OFF threshold, flow proceeds to decision diamond **442** where it is determined whether P_{background }is greater than a background threshold. If so, flow proceeds to block **452** where nonlinear processor **32** is turned off and then to block **454**. Therefore, nonlinear processor **32** is turned off when P_{error }is greater than the NLP_OFF threshold or when P_{background }is greater than the background threshold. In one embodiment, the NLP_OFF threshold is set as −27 dBm0 and the background threshold as −39 dBm0.

If it is determined at decision diamond **442** that P_{background }is not greater than the background threshold, flow proceeds to decision diamond **444** where it is determined whether P_{error }is less than a nonlinear processor on (NLP_ON) threshold. If so, flow proceeds to decision diamond **446** where it is determined whether AVG P_{Rin }is greater than a P_{Rin }threshold. If so, then flow proceeds to block **448** where NLP_ON is set (indicating that nonlinear processor **32** is turned on). Therefore, nonlinear processor **32** is turned on when P_{error }is less than the NLP_ON threshold and AVG P_{Rin }is greater than the P_{Rin }threshold. The condition of AVG P_{Rin }being greater than the P_{Rin }threshold ensures that nonlinear processor **32** is turned on only when necessary (because noticeable echo can only be the case when the far-end talker signals are relatively strong). On the other hand, the condition of P_{error }being less than the NLP_ON threshold further ensures that the residual echo has to be small and that the near-end talker signals are not mistakenly considered as residual echo to be removed. Therefore, in one embodiment, the P_{Rin }threshold is set to −36 dBm0 and the NLP_ON threshold to −42 dBm0. However, in alternate embodiments, they can be set to any appropriate value.

Note that in the embodiment described above, the different between the NLP_OFF threshold and the NLP_ON threshold (which, in one embodiment, is −15 dBm0) is a “dead zone” for nonlinear processor **32** that helps to avoid rapid switching between NLP_ON and NLP_OFF.

If it is determined that P_{error }is not less than the NLP_ON threshold (at decision diamond **444**) or the AVG P_{Rin }is not greater than the P_{Rin }threshold (at decision diamond **446**), flow proceeds to decision diamond **450** where it is determined whether NLP_ON is set (i.e. whether nonlinear processor **32** is on). If NLP_ON is not set, flow proceeds to block **232** of **448**), flow proceeds to decision diamond **456** where it is determined whether comfort noise is on. If not, flow proceeds to block **232** of **458** where comfort noise is generated. After block **458**, flow proceeds to block **460** where the comfort noise level is determined, and then to block **462** where the comfort noise is mixed with the background noise. Flow then proceeds to block **464** where the noise ramping factor is adapted and then to block **232** of

Therefore, comfort noise signals will be generated when nonlinear processor **32** is on. White noise is generally not a preferred choice for the comfort noise because it is spectrally far from the true background noise signals of everyday life. Some embodiments of the present invention therefore use pink noise, brown noise, or Hoth noise as comfort noise. For example, in one embodiment, pink noise is chosen because of its low complexity in terms of computations. A pink-like noise is generated (e.g. in block **458**) by using two consecutive realizations of uniformly distributed pseudo-random variable X as shown in equation 13 below.

*Y* _{pink}(*n*)=*C* _{1} **X*(*n*)+*C* _{2} **X*(*n−*1) Equation 13:

In equation 13 above, X(n) is the pseudo-random variable (−1≦X(n)<1) generated at time n, C_{1 }and C_{2 }are constants for modifying the mixture of the two random samples and the magnitude of Y_{pink}. Y_{pink}(n) is therefore the pink-like noise sample being generated at time n. The two constants C_{1 }and C_{2 }are chosen to ensure that the average power level of the pink noise signals is about 2 dB lower than P_{background}. For example, in one embodiment, C_{1 }and C_{2 }are chosen as 0.75 and 1, respectively. Therefore, in one embodiment, the comfort noise matching levels range from 0 to 4 dB than the estimated background noise levels.

The generated comfort noise, Y_{pink }in this embodiment, is then mixed with the background noise as shown in equation 14 below (see also block **462** of FIG. **18**).

*S*out(*n*)=α(*n*)*(error signal 46)−(1−α(*n*))**A*Y* _{pink}(*n*) Equation 14:

In equation 14 above, A is the magnitude of the background noise level to be matched (corresponding to block **460**). For example, in one embodiment, A=square root of (P_{background}). In alternate embodiments, A=P_{background}, if P_{background }is represented in magnitude, rather than power. In equation 14, α(n) is a noise ramping factor (where 0≦α<1) at time n which allows for a smooth transition from one level to another at the onset of nonlinear processor **32**, and Sout(n) is the final output of nonlinear processor **32** at time n (i.e. Sout(n) is Sout **42** of **464**) is calculated per sample as shown in equation 15.

α(*n*)=*b**α(*n−*1) Equation 15:

In equation 15, b is the ramping constant which is chosen to be less than 1. In one embodiment is approximately 0.9986 which approximately attenuates to its half in 500 ms, because 0.9986^{500}=0.496. During this ramping process, Sout(n) starts from error signal **46** (which is Sin **39**—error estimation signal **48** of _{pink}(n), as α(n) changes from 1 to 0, if the ramping process continues. The ramping can be applied on both the onset and offset of nonlinear processor **32**. However, in one embodiment, the ramping only applies to the onset of nonlinear processor **32**. The reason is that when nonlinear processor **32** is turned off, it normally detects a significant level of the near-end talker signals, and gradual switching back from the comfort noise (pink noise signals, in one embodiment) to the near-end talker signal may not be desirable. However, alternate embodiments may apply this ramping when nonlinear processor **32** is turned both on and off.

**30** which functions to estimate the pure delay. The pure delay estimation is intended for reducing the number of taps of adaptive filter **28** and thus gaining faster and deeper convergence with smaller computational effort, as was discussed above. That is, the portion of monitor and control unit **30** illustrated in **211** of **28** is reduced. Therefore,

FIGS. **7** and **20**-**24** provide one embodiment used to achieve an estimation of the pure delay (i.e. T**1** of **20**. The pure delay estimation, as will be described in more detail below, is performed to reduce the computational cost associated with covering large echo path delay spans by replacing a full-window adaptive filter with a properly positioned narrow-window adaptive filter. That is, rather than using a full-window adaptive filter covering the entire impulse response of **1** and T**4**+T**2**, a smaller window may be used (a sparse window) which excludes the pure delay portion and is positioned in order to capture T**4**+T**2**, the part during which significant responses occur. Also, the pure delay estimation increases the convergence speed and depth of adaptive filter **28** through the use of a shorter length adaptive filter. Also, the pure delay estimation may be used to monitor dynamically changing pure delay of the echo (e.g. during a phone call) and to adjust the adaptive filter window (e.g. sparse window) accordingly.

The embodiments that will be described herein may include a passive approach (e.g. sub-rate filter adaptation using the speech signal only) as well as an active approach (e.g. injecting a short, narrow-band very low level noise pulse at the beginning of the call and concurrently performing sub-rate adaptation in order to establish pure delay for calls which begin with silence on both directions, where generally, a silence lasting 300 ms is long enough to inject a low-level probing signal and determine pure delay). The embodiments to be described herein also include two scenarios for handling the pure delay. The first scenario relates to the beginning of a telephone call, where Quality of Service (QoS) principles require immediate reduction of echo. The second scenario relates to changes of the echo path in the middle of the telephone call. Typically, the sparse window (and the associated the pure delay) does not vary throughout the duration of a telephone call. However, on some calls (particularly those where, for example, ‘call forward’ or ‘conference call’ features are activated) the pure delay may change considerably. Therefore, various embodiments discussed herein support dynamics of the pure delay corresponding to up to one variation of the sparse window per second. Note that the embodiments discuss herein may use proprietary (i.e., non-standard) signaling provided via control signals **17** to determine whether a telephone is on or off hook in order to determine the beginning or end of a call.

The embodiments of **28**. The sub-rate process mentioned above may use an NLMS (Normalized Least Mean Square) adaptive filter (for adaptive filter **122** of **122** is not limited to this type of adaptive filtering. For example, PNLMS, RLS, or other adaptive filters may be used. Note that the NLMS adaptive filtering algorithm is generally simple and has acceptable convergence characteristics. Other adaptive filter algorithms are computationally more demanding. PNLMS (Proportionate Normalized LMS) algorithm offers a tangible improvement of the convergence properties at a moderate computational cost. RLS (Recursive Least Squares) adaptive algorithm is generally significantly faster (yet computational cost is also significantly greater). However, it is sensitive to numerical errors and manifests numerical instability. Other adaptive filters (such as subband, affine and their variants) may be more attractive from the viewpoint of convergence properties; in comparison with the NLMS, they are computationally more demanding though. However, the embodiments discussed herein are not limited to the use of the NLMS adaptive filtering. Both main rate adaptive filter as well as the sub-rate adaptive filter could be based on other types of adaptive filtering solutions.

The pure delay estimation may be controlled by such mechanisms as short-term sub-rate signal power estimation and sub-rate near-end talker signal detection in order to prevent generating measurements which could be inherently unreliable (as affected by either noise or near-end talk) and thus possibly causing the divergence of the sub-rate adaptive filter **122**. Note that as discussed above with reference to adaptive filter **28** where the adaptation process is stopped upon the detection of Sgen in order to avoid developing false coefficients, the same principle may apply to adaptive filter **122** used in determining pure delay.

In addition to shortening the adaptive filter length, the estimation of pure delay may be used to address other situations, such as, for example, when a far-end echo canceller is turned off, when calls are switched from local to long distance (such as via a call forward feature, call transfer feature, etc.), when conference call operations are with calling/called parties dispersed over large geographical regions, etc.

**30** used for providing estimated delay **130**. Adaptive filter **122** (which, in one embodiment utilizing sub-rate processing, is a sub-rate adaptive filter) provides a short-term estimate of the band-pass impulse response on a continuous basis, through the duration of a telephone call. The pure delay measurements of the impulse response are continuously filtered using a qualification process or decision block (e.g.

In an optional version, as will be discussed in reference to **20** may operate in a monitoring mode. In this mode, the system of **28** against a threshold and if the ERLE drops below the threshold and remains there for a predetermined duration, the system of

The flow of **482** where it is determined whether a pure delay estimation option is activated. Note that this option may correspond to a setting that is programmed into echo canceller **20**. In this case, determining whether the option is activated need not be done on a per sample basis as illustrated in **482** can be done at the beginning of a phone call. However, only if the pure delay estimation option is activated does flow proceed to decision diamond **483**. If it is not activated (whether determined at the beginning of the call or on a per sample basis), flow proceeds to block **213** of

Decision diamond **483** determines whether the optional training at the beginning of the call is activated. As with the pure delay estimation option, the training option can also be programmed into echo canceller **20** and thus checked at the beginning of a phone call rather than at each sample as illustrated in **484** of **497**. In summary, if the optional training is not activated, the flow of **20** may operate in a variety of different ways, depending on the settings and options chosen.

Also note that at the beginning of each phone call, many variables may be initialized for use in the flows of **20**. For example, the measurement cycle, N, may be initialized at the start of each call to a particular value or may be hardwired within echo canceller **20**. Note that other variables described throughout this description may be initialized at the start of a call or hardwired or programmed (permanently or not) in echo canceller **20**.

If, at decision diamond **483** it is determined that the optional training at the beginning of the call is activated, flow proceeds to decision diamond **497**. The optional training allows for a pure delay to be estimated at the beginning of a call. Since, at the very beginning of a call, there is generally no talking yet, a training signal can be injected into Rin **43** to produce Rout **40** (see training signal **41** of **43** via adder **36**). That is, in the absence of adequate Rin **43** energy, it is not possible to determine the pure delay; therefore, an injection of training signal **41** can be used to determine a pure delay estimate. Generally, training signal **41** is a short burst of relatively low energy that is injected at the beginning of a phone call, prior to a conversation. That is, training signal **41** is generally less than an injection threshold, which, in one embodiment, is in the range of −30 dBm0 to −55 dBm0. Therefore, if the optional training is activated, flow proceeds to decision diamond **497** where it is determined whether the training bypass flag is TRUE. If so, flow proceeds to decision diamond **484** of

If, at decision diamond **497**, it is determined that the training bypass flag is not set to TRUE, flow proceeds to decision diamond **499** where it is determined whether the training index is less than or equal to 2. The training index ensures that the training signal, if used, is injected only at the beginning of the call. As was mentioned above, the training index may be reset at the beginning of a call, and therefore, upon reaching decision diamond **499** for the first time, the training index should be less than or equal to 1 (since it is originally reset to zero). As will be discussed below, though, after a first measurement cycle (which, in one embodiment, is 300 milliseconds), the training index will be incremented to one (e.g. in block **505** of **505** of **499**, flow will proceed to decision diamond **484** of **41** anymore because a training index of 2 indicates that it is no longer considered the beginning of the call. It is generally undesirable to inject a training signal at another time other than the beginning of the call because it may be audible to the parties on the call.

If, at decision diamond, training index is less than or equal to one, then flow proceeds to block **489**, which indicates that it is still considered the beginning of the call. In block **489**, the long term power of Sin (P_{Sin}) and Rin (P_{Rin}) are calculated (which may be done using equations 1, 3, and 4 discussed above). Flow then proceeds to decision diamond **490** where it is determined whether P_{Sin }is less than a P_{Sin }threshold and P_{Rin }is less than a P_{Rin }threshold. The first check (whether P_{Sin }is less than the P_{Sin }threshold) ensures that there is no near end talker signal, Sgen. In one embodiment, this P_{Sin }threshold is −50 dBm0. The second check (whether P_{Rin }is less than a P_{Rin }threshold) ensures that a far-end talker signal is not present. In one embodiment, this P_{Rin }threshold is −50 dBm0. If both conditions are met, then flow proceeds to block **492** indicating that a conversation has not yet started and a training signal can be injected. Therefore, in block **492**, a training signal is injected (e.g. training signal **41** of **492**. However, if at decision diamond **490**, both conditions are not met, then flow proceeds to block **495** where the training signal flag is set to TRUE. That is, once P_{Sin }or P_{Rin }surpass their respective thresholds, training is bypassed (at decision diamond **497**) regardless of the training index, thus preventing a training signal from being injected during the current call. After blocks **495** and **492**, flow proceeds to decision diamond **484** of

**106** and **110** of ^{th }sample of Rin **44** and Sin **38** is processed. However, in alternate embodiments, D can be any value (including 1, which indicates that sub-rate processing is not used because every sample is processed). Therefore, every D-th sample is considered a sub-rate sample. A pure delay sample counter is used to keep track of the incoming samples of Rin **44** and Sin **38** in order to capture every D-th sample. Generally, the pure delay sample counter is incremented after each sample, and reset every D-th sample. The pure delay sample counter can also be reset at the beginning of each phone call, as mentioned above. Also, the delay sample counter may be shared with sample counters of other flows discussed herein, or may be a specific counter used only for estimating the pure delay.

At decision diamond **484**, it is determined whether the pure delay sample counter is equal to D-**1**. Note that in the embodiment where the pure delay sample counter is reset (i.e. set to zero), reaching D-**1** corresponds to reaching the D-th sample. However, in alternate embodiments, the pure delay sample counter may be initialized to 1 and would be checked against D rather than D-**1**. Also, other embodiments may initialize the pure delay sample counter to D or D-**1** and decrement until 1 or 0 is reached, respectively. Therefore, various embodiments of a decimation filter and decimator may be used for decimation filters **104** and **108** and decimators **106** and **110** of **106** are sub-rate samples of Rin **44**, which may be referred to as RinSR, and the output of decimator **110** are sub-rate samples of Sin **38**, which may be referred to as SinSR.

At decision diamond **484**, if it is determined that the pure delay sample counter has not yet reached D-**1**, then flow proceeds to block **502** where the pure delay sample counter is incremented by one, and flow proceeds to block **213** of **1**, flow proceeds from decision diamond **484** to block **491** indicating that a sub-rate sample has been reached. In block **491** the pure delay sample counter is reset in order to detect the next sub-rate sample as was described above.

Flow proceeds to block **485** where the power of sub-rate Rin (P_{RinSR}), the power of sub-rate Sin (P_{SinSR}), and the sub-rate near-end talker detection flag (sr_near_end_detect_flag) are determined. For example, the following equations may be used to determine P_{RinSR}, P_{SinSR}, and P_{errorSR}(k):

*P* _{RinSR}(*k*)=(1−α)·*P* _{Rin} *SR*(*k−*1)+α·*R*in*SR* ^{2}(*k*); Equation 16:

*P* _{SinSR}(*k*)=(1−α)·*P* _{SinSR}(*k−*1)+α·*S*in*SR* ^{2}(*k*); Equation 17:

*P* _{errorSR}(*k*)=(1−α)·*P* _{errorSR}(*k−*1)+α·error*SR* ^{2}(*k*); Equation 18:

Note that in the above equations (equations 16-18), k is the signal sub-rate sample number such that, for example, SinSR(k)=Sin(k·D). Equation 18 corresponds to the sub-rate error, errorSR, which corresponds to the difference between SinSR and a sub-rate echo estimate, y(k), which is determined by sub-rate adaptive filter **122** of **494**. Therefore, errorSR(k) and P_{errorSR}(k) will be described in more detail below. Also, in one embodiment of the above equations, α is set to 1/280 which corresponds to statistics of human speech as observed in a telephony channel; note that 1/280 corresponds also to approximately a 70 millisecond sliding window averaging in terms of filter bandwidth. However, alternate embodiments may use different values of alpha. (Note that the sub-rate power calculations above can be calculated by power estimator **120** and power estimator **118** of

The determination of the sr_near_end_detect_flag may be done analogously to the near-end signal detection described above with respect to _{errorSR}(k) and P_{SinSR}(k) is compared against an NESD sub-rate threshold (NESD_SR_threshold) to determine whether a near-end talker signal (Sgen) is present. (Note that this may be performed by near-end signal detector **114** of **122** because if a near-end talker signal is present, as was described above, Sin **38** is no longer representative of the pure residual echo but instead is a mixture of both Sgen and the residual echo. Therefore, as discussed above with reference to adaptive filter **28**, sub-rate adaptive filter **122** should adapt only when SinSR includes only the sub-rate echo (i.e. when a near-end talker signal is not present). Also, as discussed above with reference to adaptive filter **28**, sub-rate adaptive filter **122** should adapt when P_{RinSR }is sufficiently high to prevent adaptation to channel noise.

Note that as above with reference to adaptive filter **28**, a near-end talker signal can be detected during both a single talk and a double talk situation. That is, using the above algorithm, Sgen can be detected when only a near-end talker is present or when both a near-end and a far-end talker are present. Also note that alternate embodiments may use other methods for determining if a near-end talker signal is present. For example, one embodiment may use a Geigel algorithm, which is known in the art to detect a near-end talker signal.

After block **485**, flow proceeds to decision diamond **486** where it is determined whether P_{RinSR }is greater than a minimum power threshold of sub-rate Rin. If not, then flow proceeds to block **213** of **122**. As mentioned above, this prevents sub-rate adaptive filter **122** from adapting to channel noise. In one embodiment the minimum power threshold of sub-rate Rin is set to −45 dBm0. If the minimum threshold is met, flow proceeds to decision diamond **487** where it is determined whether the sr_near_end_detect_flag is FALSE. If minimum threshold is not met, then flow proceeds to block **213** of **122** due to the presence of a near-end talker signal, as discussed in the previous paragraph. If the sr_near_end_detect_flag is FALSE, flow proceeds to block **494**, which indicates that P_{RinSR }is sufficient and no near-end talker signal is present.

In block **494**, the sub-rate echo estimate, y(k), is calculated, and then in block **496**, the coefficients of sub-rate adaptive filter **122** are updated. In one embodiment, a modified NLMS algorithm (modified for use in a sub-rate process) may be used to calculate y(k) and update the coefficients.

*y*(*k*)=*X* ^{T}(*k*)·*H*(*k*) Equation 19:

Equation 19 above represents FIR filtering of the input signal X where X(k)=[x(k), x(k−1), . . . , x(k−N+1)]^{T }is the input signal vector (at the sub-rate D) extending over the duration of the FIR filter span. Therefore, x(n)=RinSR(n). Also, in equation 19, H(k) is filter coefficient vector (at the sub-rate sampling) for the k-th iteration where:

*H*(*k*)=[*h* _{0}(*k*), *h* _{1}(*k*), . . . , *h* _{N−1}(*k*)]^{T} Equation 20:

*H*(*k+*1)=*H*(*k*)+step_size·error*SR*(*k*)·*X*(*k*) Equation 21:

Equation 21 above represents the filter coefficients update formula, as per an NLMS algorithm, where the NLMS sub-rate step_size can be expressed as follows.

step_size=β/[γ+*P* _{RinSR}(*k*)] Equation 22:

In equation 22, β is an adaptation constant and γ is a ‘protection’ term, which ensures that the update term in the adaptation formula does not become excessively large when P_{RinSR}(k) temporarily becomes small, and where P_{RinSR}(k) is the input signal power at the sub-rate sampling (see equation 16).

error*SR*(*k*)=*S*inSR(*k*)−*y*(*k*) (adaptation error at the sub-rate) Equation 23:

In the above equations, RinSR corresponds to the filtered and decimated far-end signal (which corresponds to the output of decimator **106** of **110** of **38** includes only the residual echo, and therefore SinSR at the output of decimator **110** can be used as the filtered and decimated echo signal. The variable H (discussed in reference to equation **19**) corresponds to a column vector representing the sub-rate adaptive filter **122** coefficient estimates, and the “T” following the H indicates a vector transposition. The signal y represents the estimate of SinSR provided by adaptive filter **122**, and errorSR is the difference between SinSR and y. Also, in one embodiment of the above equations, β is set to 2^{−9}*2.5, and α to 1/128. In one embodiment, γ is set to a small value (comparing to P_{RinSR}(k) ). For example, if P_{RinSR}(k) is represented as a 16-bit fractional number, a typical value for γ is k·2^{−15}, where k is a small integer.

Flow then proceeds to decision diamond **498** where it is determined whether n is equal to N, where in the current embodiment, N corresponds to the duration of a single measurement cycle. In one embodiment, N corresponds to 300 ms, and therefore, corresponds to 300 sub-rate samples (assuming D=8). For example, if the signals (e.g. Rin **44** and Sin **38**) are being sampled at a rate of 8 kHz, then a sample is received every 125 microseconds. In the current example, D is 8; therefore, every D-th sample corresponds to 8*125 microseconds, which equals 1 millisecond. Therefore, every N sub-rate samples, flow proceeds to blocks **503**, **500** and **501** where, in the current embodiment, N is 300 such that 300*1 millisecond is 300 milliseconds. Therefore, N can be defined as either a time window having a predetermined duration or as a predetermined number of sub-rate samples that must be processed prior to determining the estimated delay in blocks **500** and **501**. The value for N may be programmed or hardwired within echo canceller **20**, and may be any value, depending on the desired frequency of calculating a new estimated delay value. Note that N corresponds to the convergence time (i.e. a short-term convergence time) for sub-rate adaptive filter **122**. For example, if a window of 1024 samples (which, in the current embodiment, corresponds to 1024*125 microseconds which equals a window size of 128 milliseconds, assuming a base sampling rate of 8 kHz) is used to capture the impulse response (such as T**3** in

If, at decision diamond **498**, it is determined that the index n (which may be initialized at the beginning of the call to a starting value such as 1 or 0, for example) has not yet reached N, flow proceeds to block **502** where n is incremented, and flow continues to block **213** of **498**, it is determined that n is equal to N, indicating that 300 samples have been processed (corresponding to a duration of 300 milliseconds), flow proceeds to block **503** where n is initialized to 1, and other measurement cycle variable are also initialized (e.g. P_{RinSR}, P_{SinSR}, sr_near_end_detect_flag, etc.). Flow then proceeds to decision diamond **504** where it is determined whether the training index is 2. If so, flow proceeds to block **500**, bypassing block **505**. However, if training index is not 2, flow proceeds to block **505** where the training index is incremented. As described above in reference to

Flow proceeds from block **505** or decision diamond **504** to block **500** where the individual estimated pure delay is calculated. Note that, as will be described in more detail in reference to **501** where, using several (2, 3 or more, depending on a particular implementation as well as depending on the stage of the call) valid individual pure delay estimations, an estimated delay **130** is determined, as will be discussed in more detail in reference to

**500** of **506** where the sub-rate echo return loss enhancement (SR_ERLE) is determined. The following equation may be used to determine the SR_ERLE:

*SR* _{—} *ERLE*(*k*)=10*log_{10}(*P* _{SinSR}(*k*)/P_{errorSR}(*k*)) Equation 24:

The SR_ERLE therefore corresponds to a ratio between P_{SinSR }and P_{errorSR}, which is used for validating the pure delay measurements. SR_ERLE provides information on the “goodness” of the convergence of sub-rate adaptive filter **122** (i.e. how much echo was cancelled). That is, a higher SR_ERLE (such as, for example, 5 dB or more) indicates that within the current measurement cycle, adaptive filter **122** has converged sufficiently. (Note that SR_ERLE can be determined by ERLE estimator **116** operating at the given sub-rate, see **506**, flow proceeds to decision diamond **508** where SR_ERLE is compared to a sub-rate ERLE threshold, and if it not greater than this threshold, flow proceeds to block **514**, indicating that the current measurement cycle should not be used due to its poor SR_ERLE. Therefore, the current measurement (for the current measurement cycle) is discarded and flow proceeds to block **501** of **510** which performs another check on the convergence of sub-rate adaptive filter **122**.

In block **510**, the peak-to-average ratio (PAR) of sub-rate adaptive filter **122** coefficients is determined. Referring to **512**, flow proceeds to block **514**, where the current measurement is discarded because the current measurement cycle did not provide for adequate convergence of sub-rate adaptive filter **122**. However, if the PAR is greater than the PAR_Threshold, flow proceeds to block **516**, indicating that two conditions were met to ensure that sub-rate adaptive filter sufficiently converged during the current measurement cycle. In block **516**, the maximum value of the sub-rate adaptive filter **122** coefficients (corresponding to the peak) is located (which may be performed by maximum value locator **124** of **501** of

**501** of **126** and estimated delay **130** of **528**, where it is determined whether this is the first time through the flow (i.e. indicating the beginning of a phone call) or if the previous estimated delay was equal to zero (which may correspond to either the beginning of a call or the middle of call having a previous estimated delay value of zero). If either one of these cases is true, flow proceeds to decision diamond **529** where it is determined if two valid measurements are available. As discussed above in reference to **530** where a “fast track” calculation of the estimated delay begins with block **530**.

In block **530**, a first buffer is filled with two consecutive valid measurements. Flow proceeds to block **532** where the dispersion between these two measurements and the average of the two measurements are taken. The dispersion, for example, can be the difference between the two measurements. Flow proceeds to decision diamond **534** where it is determined whether the dispersion is less than a dispersion threshold **1** and the average is greater than an average threshold **1**. If not, a new estimated delay is not calculated and flow proceeds to block **213** of **542** where a new estimated delay is calculated. Therefore, dispersion threshold **1** and average threshold **1** ensure that a new estimated delay is calculated only if the two measurements are consistent enough with each other. In other words, if the subsequent measurements of the time for which the impulse response reaches its maximum value differ too much, the estimate of the delay is put on hold until the subsequent measurements are more consistent (i.e., closer to each other). In block **542**, the following equation can be used to calculate the new estimated delay:

new estimated delay=average−offset Equation 25

In the above equation, the average corresponds to the average taken in block **532** of the two measurements, and the offset is the value corresponding to the amount of time before reaching the peak within the impulse response that a substantial response began. That is, referring to **1** (the pure delay) by an amount of T**4**. Therefore, T**4** must be subtracted from the value of the time at the peak (Tpeak). The offset corresponds to this T**4** value, and the offset can be determined using statistical information about impulse responses of different yet typical hybrid circuits present in the field and can be programmed in echo canceller **20**. The new estimated delay (corresponding to estimated delay **130**) is then applied in block **544**. For example, applying estimated delay **130** may correspond to enabling the optional delay block **66** in **28**. Therefore, through the use of the pure delay estimation, the number of filter taps required by adaptive filter unit **28** is reduced because the coefficients for the pure delay portion of the response can be considered to be zero.

Note that alternate embodiments may require more or less than two measurements in decision diamond **529** to continue with the “fast track” calculation. In one embodiment, only one valid measurement may be required, and in this case, the dispersion and average are not calculated (since only one measurement is used). Also, the actual value can therefore be checked against the average threshold **1** before determining whether to proceed to block **542**, and the dispersion comparison would not be needed. In an alternate embodiment where more than two valid measurements are required, the dispersion may correspond to a variance taken with respect to the valid measurements. Therefore, alternate embodiments may require any number of valid measurements.

If, at decision diamond **528**, it is determined that this is not the first time through the call (i.e. generally indicating that the estimation of pure delay is performed in the middle of the call) and the previous delay estimate was not zero, flow proceeds to decision diamond **535** where it is determined whether M valid measurements are available. In one embodiment, M is selected to be 3, or 4, or 5 (depending on the particular setting chosen by the echo canceller installer). The value of M may be chosen such that more or less valid measurements are required before the possibility of updating (i.e. changing) the current estimated delay value. The higher the M value, the less often flow will proceed to block **536**. Therefore, M may be chosen to be any value and is not limited to 3 through 5. If M valid measurements are not available, flow proceeds to block **213** of **536** where a second buffer is filled with M consecutive valid measurements.

In block **538**, a dispersion, average, and a difference between the average and a previous average are calculated. As described above, the dispersion may be calculated in a variety of ways. For example, if M is only 2, the dispersion can simply be a difference. Alternatively, the dispersion can be calculated as a variance. The previous average corresponds to the average calculated in the previous pass through either block **538** or **532**. After the calculations of block **538**, flow proceeds to decision diamond **540** where various thresholds are used to determine whether a change in the estimated delay is worth while. Therefore, the thresholds of decision diamond **540** can be used to set up more conservative criteria for changing the estimated pure delay in the middle of a call.

At decision diamond **540**, the dispersion is compared to a dispersion threshold **2**, the average to an average threshold **2**, and the difference between the average and the previous average to a difference threshold. If the dispersion is less than the dispersion threshold **2**, or the average is greater than the average threshold **2**, or if the difference is less than the difference threshold, flow proceeds to block **213** of **2**, average is greater than average threshold **2**, and the difference is greater than the difference threshold), then flow proceeds to block **542** where a new estimated delay is calculated (as explained in reference to equation 25) and applied in block **544**, as described above. As with the “fast track”, the dispersion threshold **2** ensures that the M valid measurements do not deviate from each other too much and the average ensures that the M valid measurements are large enough to warrant the necessity of changing it (for example, if the average is relatively small, there may not be a need to change the pure delay of the echo canceller, as the small pure delay can be accommodated by the adaptive filter **28** if properly provisioned). The comparison of the difference to a difference threshold prevents the current estimated delay from being changed if the difference is too small (i.e. less than the difference threshold) and therefore not worth changing.

**20**. The flow of **211** of **518** where the echo return loss enhancement (ERLE) is calculated. This ERLE corresponds to the “goodness” of the convergence of adaptive filter **28** (i.e. provides information as to how much echo was actually not cancelled out by adaptive filter **28**). The following equation may be used to calculate ERLE:

*ERLE*(*n*)=10*log_{10}(*P* _{Sin}(*n*)/*P* _{error}(*n*)) Equation 26:

The ERLE therefore corresponds to a ratio between P_{Sin }and P_{error }where n is the sample number. (Note that P_{Sin }and P_{error }can be calculated using equations 1 and 2 described above.) This ERLE value is therefore used during the monitoring mode for entering the pure delay adjustment process of **520** where ERLE is compared against an ERLE threshold. If ERLE greater than or equal to the ERLE threshold, the convergence of adaptive filter **28** is sufficient and the pure delay estimation need not be performed; therefore, flow proceeds to block **213** of **28** is not sufficient, and flow proceeds to block **521** where an ERLE counter is incremented. (Note that this ERLE counter can be initialized at the beginning of each call.). Flow then proceeds to decision diamond **523** where the ERLE counter is compared to an ERLE counter threshold. If the ERLE counter has not reached the ERLE counter threshold, flow bypasses block **522** (corresponding to the flow of **213** of **522** where the entire flow of **524** where the ERLE counter is reset, and then to block **213** of

The ERLE counter and ERLE counter threshold ensure that if the ERLE calculated in block **518** is borderline (changes from above the ERLE threshold to below the ERLE threshold occur too frequently), the pure delay does not get recalculated and updated. That is, the ERLE has to fall below the ERLE threshold for a period of time (controlled by the ERLE counter and ERLE counter threshold) before the flow of

FIGS. **8** and **25**-**27** relate to one embodiment of tone detection that may be used within echo canceller **20**, where **30**, and **209** of **20** (e.g. Rin **44** or Sin **38**) is a tone, stability of adaptive filter **62** may be affected, resulting in undesirable distortion and degraded quality of service in telecommunication networks. A tone is a signal composed of a number of sinusoidal components with constant magnitude, frequency and phase over a certain period of time.

Any adaptive algorithm (such as that used by adaptive filter **62**) attempting to minimize the average power of the residual echo will have a dynamic behavior that depends on the auto-correlation matrix of Rin **44**. Certain classes of receiving path signals can make this matrix singular, which can temporarily disrupt the adaptation process and make the filter coefficients of adaptive filter **62** deviate from desirable values. Sinusoidal signals (single-frequency tones), for example, can create this condition. In this case, the auto-correlation, r(k), of a sinusoidal signal Rin(n)=A cos(Ωn+φ) is given by r(k)=A^{2 }cos(Ωk)/2, which leads to a singular auto-correlation matrix in most practical cases (i.e. when its dimension is large). When that happens, a possible outcome of the adaptive algorithm is a set of filter coefficients (for adaptive filter **62**) with sinusoidal form, which is an incorrect estimate of the actual hybrid circuit impulse response, an example of which is given in

Similarly, multi-frequency tones can also generate a similar problem because their auto-correlation

can still generate singular auto-correlation matrices when the number of components, M, is not large enough or the matrices have a large dimension. Note that the matrix dimension depends on the number of filter coefficients being used to estimate the impulse response of the hybrid circuit. Therefore, it is desirable to detect the presence of any signaling and controlling tones and then to stop the adaptation process of adaptive filter **62**, thus preventing divergence from a good set of filter coefficients.

One embodiment which will be described in reference to **20**, they may be used in any device or telecommunication device that requires tone indication and detection, and is not limited to echo cancellers alone.

**134** which maps any single-frequency tone to a constant via a modified energy operator. That is, a single-frequency tone can be expressed as follows.

*x*(*n*)=*A *cos(Ω*n*+φ) Equation 27:

The modified energy operator, Ψ_{k}, can be expressed as follows.

Ψ_{k}(*x*(*n*))=*x* ^{2}(*n−k*)−*x*(*n*)*x*(*n−*2*k*)=A^{2 }sin^{2}(*k*Ω) Equation 28:

In the above equation, note that x^{2}(n−k)−x(n)x(n−2k) corresponds to the output of adder **144** in **136** is x(n−k), the output of delay **138** is x(n−2k), the output of multiplier **140** is x(n)x(n−2k), the output of multiplier **142** is x^{2}(n−k), and the output of adder **144** is the sum of the output of multiplier **142** and the negative of the output of multiplier **140**). Note that the input signal x(n) can correspond to either Rin **44** or Sin **38**. Furthermore, by substituting x(n) of equation **27** into x^{2}(n−k)−x(n)x(n−2k), the result A^{2 }sin^{2 }(kΩ) is obtained. Therefore, note that Ψ depends both on the magnitude A and the normalized frequency Ω of the tone (Ω=2πf/f_{s}, where f is the tone frequency and f_{s }is the sampling frequency, which, in one embodiment, is 8 kHz). The parameter k in these equations defines the underlying sub-rate processing, where k can be any integer value, including 1. Therefore, applying Ψ_{k }at a sampling rate f_{s }is equivalent to applying Ψ_{1 }at a sampling rate of f_{s}/k. As described above, sub-rate processing may be used to reduce computational requirements, where only every kth sample is processed. Note also that Ψ_{k}(x(n)) does not depend on the initial phase φ, but does generate a short-term transient upon abrupt phase changes, which maybe used to detect phase changes in the communication signal x(n).

The power of x(n) (equation 27) can be expressed using the following equation.

Power_{x(n)} *=A* ^{2}/2 Equation 29:

Therefore, note that Ψ_{k}(x(n)) provides the power of x(n) scaled by 2 sin^{2}(kΩ), such that:

Ψ_{k}(*x*(*n*))=Power_{x(n)}*2 sin^{2}(*k*Ω) Equation 30:

Solving for Power_{x(n) }in terms of Ψk(x(n)) therefore provides the following equation:

Power_{x(n)}=Ψ_{k}(*x*(*n*))*csc* ^{2}(*kΩ*)/2 Equation 31:

However, in practice, the signal x(n) (which, as mentioned above, may correspond to either Rin **44** or Sin **38** in the embodiment illustrated in _{k}(x(n)). Any low pass filter can then be used to smoothen the result, such as, for example, a single-pole low pass filter. Therefore, as can be seen in **134** includes a low-pass filter which receives the output of magnitude **146** (corresponding to the absolute value of the output of adder **144**) and a from storage **150**, and provides a smooth estimate P(n) of Ψnoisy_{k}(x(n)). P(n) can be expressed with the following equation.

*P*(*n*)=*α P*(*n−*1)+(1−α)|*x* ^{2}(*n−k*)−*x*(*n*)*x*(*n−*2*k*)| Equation 32:

In the above equation, α is a smoothing parameter (0<α<1) that controls the bandwidth of the smoothing low pass filter. Note that either a fixed or variable smoothing parameter, α, may be used. P(n) is then provided to tone indication decision unit **166** of

Once a tone is present, one embodiment detects any pre-defined single-frequency tone, with or without phase reversal, such as a 2100 Hz signaling tone. One embodiment for detecting the pre-defined tone will be discussed in more detail below with reference to **62** after a tone is received. That is, the transition between signaling tones and voice signals can also be detected using the variance of P(n) (i.e. the transition being detected when the variance is larger than some pre-defined threshold).

Given estimates P(n), tone indication decision unit **166** may be used to detect a tone according to the flow of **588** where k, a, m, r, P_{low}, and N_{min }are set to desirable values. Depending on the expected tone frequency range and the noise level in the system, those values could be, for example, k=2, a=0.9, m=1, r=0.95, P_{low}=2^{−8}. N_{min }depends on the sampling rate and the minimum required duration of the tone to be detected. Flow proceeds to decision diamond **590** where it is determined whether P(n) is greater than P_{low}, where P_{low }corresponds to a threshold indicating the lowest signal level to be considered. If not, flow proceeds to block **598** where a detection counter is reset (to zero) and then to block **604**, indicating a tone is not detected, and then to block **554** of _{low}, flow proceeds from decision diamond **590** to block **592** where P_{min }and P_{max }are computed. P_{min }corresponds to the minimum of two estimates of P(n) separated by m samples, and P_{max }corresponds to the maximum of two estimates of P(n) separated by m samples.

*P* _{min}=MIN(*P*(*n*), *P*(*n−m*)) Equation 33:

*P* _{max}=MAX(*P*(*n*), *P*(*n−m*)) Equation 34:

The variance level is estimated by comparing P_{min }and P_{max}. Therefore, flow proceeds to decision diamond **594** where the ratio of P_{min }to P_{max }(i.e. P_{min}/P_{max}) is compared to a tone indication threshold. If it is not greater than the tone indication threshold r, flow proceeds to block **598** where the detection counter is reset, then to block **604** indicating that a tone is not detected, and then to block **554** of _{min}/P_{max }is greater than the tone indication threshold, then P(n) is considered sufficiently constant (i.e. P_{min }and P_{max }are close enough) indicate the possibility of the presence of a tone. In this case, flow proceeds to block **596** where the detection counter is incremented (note that the detection counter can be initialized or reset at the beginning of a call or at any other appropriate time prior to entering the flow of

Flow proceeds to decision diamond **600** where it is determined whether the detection counter is greater than N_{min}. If the detection counter has not reached N_{min}, then flow proceeds to block **604** indicating that a tone was not detected, and then to block **554** of _{min}, flow proceeds to block **602** where a tone is detected (which corresponds with the assertion of tone indicator signal **168** in **554** of _{low}), the variance of P(n) is smaller than a minimum value (related to the tone indication threshold), and the detection counter is larger than a minimum value (N_{min}). The detection counter ensures that a tone has been present for at least a predetermined amount of time (corresponding to N_{min}) before detecting a tone and asserting tone indicator signal **168**. This helps to prevent rapid switching between detecting a tone and not detecting a tone which may result in enabling or disabling the adaptive process of adaptive filter **62** too frequently.

**152**. This correlator can be used in a variety of ways, including detection of any predefined single frequency tone, detection of a carrier of amplitude-modulated signals, detection of multi-component tones whose frequencies are close to a nominal frequency. Smooth correlator **152** receives samples of the input signal x(n) (which, as mentioned above, can be Rin **44** or Sin **38**) and three control parameters (c, b, and e) from storage **150** and generates two correlation estimates R_{0}(n) and R_{1}(n). These correlations are used to indicate the presence of a predefined tone, as will be explained as follows. The control parameter c defines one of the coefficients of a second order digital oscillator w(n) that generates a pre-defined single-frequency tone with normalized frequency Ω_{d}=2πf_{d}/f_{s}, where, as above, f_{s }is the input sampling frequency. The oscillator is initialized with the states w(−1)=1 and w(−2)=c=cos(Ω_{d}), and uses the standard second order digital oscillator given by w(n)=2*c*w(n−1)−w(n−2). (Note that the oscillator may correspond to oscillator **164** of **156** and **158**.) The input signal x(n) and a delayed version x(n−e) (i.e. the output of delay **154** of **156** and **158**) and then passed through low-pass filters (i.e. low-pass filter **160** of **158** which can be represented as x(n)w(n), and low-pass filter **162** of **156** which can be represented as x(n−e)w(n)). The parameter, b, provided as an input to low-pass filters **160** and **162** where 0<b<1 defines the bandwidth of the low-pass filters. Also, one embodiment of smooth correlator **152** may use smoothing single pole low-pass filters for low-pass filters **160** and **162**. Also, in an alternate embodiment, the oscillator signal w(n) and a delayed version w(n−e) may be correlated with x(n) rather than correlating w(n) to x(n) and x(n−e). Also, in one embodiment, e is a delay factor expressed as follows.

The above equation corresponds to a phase difference close to 90°, where ┌x┐ indicates the smallest integer larger than or equal to its argument, x.

Referring back to **160** is correlation estimate R_{0}(n) and the output of low-pass filter **162** is correlation estimate R_{1}(n), both of which are provided to tone indication decision unit **166**. R_{0}(n) and R_{1}(n) can be expressed as R_{0}(n)=b·R_{0}(n−1)+(1−b)·w(n)·x(n), and R_{1}(n)=b·R_{1}(n−1)+(1−b)·w(n)·x(n−e). Therefore, given that an unknown tone has been indicated by tone indication decision unit (using the flow of _{0}(n) and R_{1}(n) are analyzed using the flow of **164**).

The flow of **606** where c, e, b, u, q, and M_{min }are set to desirable values. The parameter c is directly related to the frequency of the target tone to be detected, which also defines the delay value e. Depending on the noise level in the system, the remaining values could be, for example, b=0.9, u=1, q=0.95, P_{low}=2^{−8}. M_{min }depends on the sampling rate and the minimum required duration of the target tone to be detected. Flow then proceeds to block **608** where R(n), R_{min}, and R_{max }are evaluated, as shown in the following equations.

*R*(*n*)=MAX(|*R* _{0}(*n*)|, |*R* _{1}(*n*)|}) Equation 36:

*R* _{min}=MIN(*R*(*n*), *R*(*n−u*)) Equation 37:

*R* _{max}=MAX(*R*(*n*), *R*(*n−u*)) Equation 38:

R(n) refers to the peak magnitude correlation between R_{0}(n) and R_{1}(n). R_{min }corresponds to the minimum of two estimates of R(n) separated by u samples, and R_{max }corresponds to the maximum of two estimates of R(n) separated by u samples. Flow then proceeds to decision diamond **610** where the ratio of R_{min }to R_{max }(R_{min}/R_{max}) is compared to a correlation threshold q (which, in one embodiment, is set to 0.95). If the ratio is not greater than the correlation threshold, flow proceeds to block **616** where the correlation counter is reset (to zero), then to block **618** indicating that the pre-defined frequency is not detected (D_{r}=0), and then to block **560** o **612** where the correlation counter is incremented. (Note that the correlation counter can also be initialized or reset at the beginning of a call.) Flow proceeds to decision diamond **614** where the correlation counter is checked against M_{min}. If the correlation counter is not greater than M_{min}, flow proceeds to block **618** indicating that the pre-defined frequency is not detected (D_{r}=0).

However, if at decision diamond **614**, it is determined that the correlation counter is greater than M_{min}, flow proceeds to decision diamond **620** where it is determined if R(n) is equal to the absolute value of R_{0}(n). If so, flow proceeds to block **622** where the pre-defined frequency is detected with a sign of R_{0}(n), i.e. D_{r}=sign(R_{0}(n)). If not, flow proceeds to block **624** where the pre-defined frequency is detected with a sign of R_{1}(n), i.e. D_{r}=sign(R_{1}(n)). From block **622** or **624**, flow proceeds to block **560** of _{min}. This helps prevent rapid switching, as described above with respect to the detection counter and N_{min}. The method of

*R* _{eff}(*n*)=½*{[R* _{0}(*n*)−*R* _{1}(*n*)] sign(|*R* _{0}(*n*)|−|*R* _{1}(*n*)|)+[*R* _{0}(*n*)+R_{1}(*n*)]} Equation 39:

The above equation generates either R_{0}(n) or R_{1}(n) depending on the component with the largest magnitude.

One embodiment of an overall process flow including the flows of **209** of **550** where minimum counter values (L_{min-p }and L_{min-n}) are selected for D_{positive }(D_{p}) and D_{negative }(D_{n}), respectively. These values are selected such that desirable minimum durations of positive and negative phases are met. D_{p }corresponds to a counter for positive phase and D_{n }for negative phase.

Flow then proceeds to block **552** where a search for any single-frequency tone is detected. The flow of **554**, where, if no tone is detected, flow proceeds to block **558** where the D_{p }and D_{n }counters are reset (to zero) and flow proceeds to block **582**, indicating that a tone is not detected, and then to block **211** of **556** where the detected tone is correlated with a pre-defined single-frequency tone. Therefore, the flow of **556**. Flow proceeds to decision diamond **560** where it is determined whether D_{r }is zero. If so, the pre-defined frequency was not determined in block **556** (e.g. block **618** of **558** where counters D_{p }and D_{n }are reset. However, if not, flow proceeds to decision diamond **562** where it is determined whether D_{r }is greater than zero. If so, flow proceeds to block **564** where the positive phase counter is incremented; otherwise, flow proceeds to block **566** where the negative phase counter is incremented.

After block **564** or **566**, flow proceeds, via point G, to block **568** where Flag_{positive }(F_{p}) and Flag_{negative }(F_{n}) are reset (to zero). Flow proceeds to decision diamond **570** where, if D_{p }is greater than L_{min-p}, F_{p }is set to one in block **572**, otherwise flow proceeds to decision diamond **574**, bypassing block **572**. At decision diamond **574**, it is determined whether D_{n }is greater than L_{min-n}, and if so, F_{n }is set to one in block **576**. If not, flow proceeds to decision diamond **578**, bypassing block **576**. At decision diamond **578**, if F_{p }and F_{n }are zero (i.e. if F_{p}+F_{n }is zero), flow proceeds to block **582** indicating that a tone was not detected. That is, if none of the counters (D_{n }or D_{p}) are larger than some minimum value (e.g. L_{min-p }or L_{min-n}, respectively), the desired tone is not detected.

However, if Fp+Fn is not zero, flow proceeds to decision diamond **580** indicating that a tone is detected. If only one counter is larger than L_{min }(D_{p }or D_{n}) then F_{p}+F_{n }is not greater than one, and flow proceeds to block **584** indicating that the desirable tone is detected without correlation sign reversal. If both D_{p }and D_{n }are larger than their respective L_{min}, flow proceeds to block **586** indicating that the desirable tone is detected with correlation sign reversal. If the average level of R(n) is the same during a correlation sign reversal, then a phase reversal is indicated. Therefore, the flow of

Note that the description up to this point has assumed that optional non-adaptive filter **64** within adaptive filter unit **28** was not present (see **28** was analogous to referring to the coefficients or taps of adaptive filter **62** within adaptive filter unit **28**. Therefore, in the previous descriptions, it was not necessary to refer to adaptive filter **62** separately from adaptive filter unit **28**. However, in the descriptions to follow of **64** may be present and may be considered as a portion of adaptive filter unit **28**. Therefore, the coefficients of adaptive filter **62**, as was used throughout the above descriptions, will now be more specifically referred to as the coefficients or taps of adaptive filter **62** since adaptive filter unit **28** may include a combination of various different filters such as adaptive filter **62** and non-adaptive filter **64**.

As described above, adaptive filter **62** tracks the echo introduced by hybrid **16**, thus generally requiring a large number of taps. For example, in order to track the entire impulse response of **62** of adaptive filter unit **28** (assuming a sampling rate of 8 kHz) requires 256 taps which span 32 milliseconds, thus covering the entire impulse response. As the number of taps of adaptive filter **62** increases, computation complexity increases and usually degrades the speed of convergence. The methods described above with respect to **62** to use a sparse window covering the impulse response after the pure delay. The methods that will be described below in reference to

Intrinsic to the impulse response, as illustrated in **31** and **35** such that the output of non-adaptive filter **31** (i.e. Sin **39**) is equivalent to the echo convolved with non-adaptive filter **31**. However, the presence of non-adaptive filter **31** following DC notch filter **45** introduces distortion into Sin **37** which needs to be compensated for. Therefore, non-adaptive filter **35** can be introduced to receive error signal **46** and produce filtered error signal **47**. Assuming non-adaptive filter **31** is an FIR filter having a transfer function A′(z), non-adaptive filter **35** is an inverse IIR filter having a transfer function 1/A′(z). However, restrictions are needed on A′(z) because the zeroes of A′(z) of FIR filter **31** become poles of 1/A′(z) of inverse IIR filter **35**. These restrictions on the zeroes of A′(z) will be described further below and prevent the poles of non-adaptive IIR filter **35** from blowing up the error signal **47**.

An alternate embodiment may use a different arrangement of the non-adaptive filters. For example, non-adaptive filter **35**, rather than being placed at the output of adder **34**, can be placed prior to adder **34** at the output of both non-adaptive filter **31** and adaptive filter **62** (which would result in the same net effect). In this embodiment, non-adaptive filter **31** followed immediately by non-adaptive filter **35** would effectively cancel each other out, so as to require no filter between Sin **38** and Sin **39** (i.e. Sin **39** and Sin **38** would be equivalent). Adaptive filter **28** can then be designed to include optional non-adaptive filter **64** (analogous to the non-adaptive filter **35**). Therefore, in this embodiment, only one additional filter is needed. However, if an IIR filter is used for non-adaptive filter **64**, restrictions on stability are still required. That is, all roots of the polynomial defined by the coefficients of filter **64** should be less than one (i.e. within a unit circle), as will be described in more detail below. Note that as used throughout the description, for a transfer function H(z)=W, the roots of W correspond to the zeroes of H(z), while for H(z)=1/W, the roots of W correspond to the poles of H(z). The optional filters **31**, **35** and **64** are non-adaptive in the sense that their coefficients are not periodically adapted as the coefficients of the main adaptive filter **62**. In general, they can be viewed as adaptive filters whose adaptive rates are event driven.

**213** of **626** where it is determined whether the adaptive filter shortening estimation option is enabled. If not, flow bypasses the flow of **212** of **628**. The adaptive filter shortening estimation option can be enabled in a variety of different ways. For example, it can be self enabled, such as in response to a system reset. Alternatively, it may be enabled any time a different delay is detected within delay unit **66** (if present) of **64** or non-adaptive filters **31** and **35** depend on the particular hybrid **16** because each different hybrid may have a different impulse response with different pure delay or different dispersion times.

In one embodiment, the pure delay estimation described above in reference to **64** or non-adaptive filters **31** and **35**. The pure delay calculation in **62**. The method of **66** to compensate for the added filter (**64** or filters **31** and **35**). That is, as will be described below, the addition of filters to shorten the dispersion time also tends to slightly increase the pure delay amount, and therefore, the delay of delay unit **66** (originally determined by the method of **66**. Furthermore, the adaptive filter shortening estimation option of **66** by the flow of

In alternate embodiments, the flow of **62** as well as the coefficients of the additional non-adaptive filter **64** or additional non-adaptive filters **31** and **35**), the additional filter or filters may simply be bypassed (or they may pass the signal through unfiltered).

If the option is enabled at decision diamond **626**, flow proceeds to decision diamond **628** where it is determined whether ERLE is good enough. (The ERLE can be calculated as shown above in equation **26** where the ERLE corresponds to a ratio between P_{Sin }and P_{error}, and where P_{Sin }and P_{error }can be calculated using equations 1 and 2 described above.) To determined if ERLE is good enough, it can be compared to a threshold. For example, the threshold can be set to a value of greater than 20 dB, or alternatively, can be set within a range of 30 to 40 dB. Generally, the higher the ERLE, the better the signal (because the lower the error, P_{error}). A high enough ERLE occurs when no Sgen (near-end talker signal) is present, because otherwise, the ERLE drops. Alternatively, it can be determined from near-end signal detector **26** described above whether a near-end talker signal exists before continuing to determine the ERLE and comparing it against a threshold. If a near-end talker signal exists, or if ERLE is not good enough, flow continues to block **212** of **630**. That is, the adaptive filter shortening estimation option should be performed when a good signal is present and error signal **46** is obtained from a good echo estimation **48**. Note that in alternate embodiments, many signals within the system may be used to determine whether good signals are present for performing the option.

In block **630**, the pure delay and dispersion based on the current coefficients of adaptive filter **62** are determined. (Note that the details of block **630** will be described in reference to **630**, flow proceeds to block **632** where, based on the pure delay and dispersion determined in block **630**, the coefficients, W, of the additional non-adaptive filter **64** or filters **31** and **45** as well as the new coefficients of adaptive filter **62** corresponding to the new shortened version of adaptive filter **62** are determined. (Note that the details of block **632** will be described in reference to **634** where it is determined if the new configuration is good enough. That is, different criteria may be used to determine whether the new configuration is satisfactory. For example, in one embodiment, if the reduced number of coefficients of the new configuration is still greater than the desired number of reduced coefficients, the process of block **632** can be repeated in an effort to obtain a further reduced number of coefficients. Alternatively, decision diamond **634** may not be present, such that only one iteration is performed, and the result of block **632** is considered sufficient.

If the new configuration is good enough at decision diamond **634**, flow continues to block **636** where adaptive filter **62** is reconfigured. That is, the new coefficients for adaptive filter **62** determined in block **632** are loaded into adaptive filter **62** and adaptive filter **62** is used in combination with non-adaptive filter **64** or in combination with non-adaptive filters **31** and **35**. (Note that the details of block **636** will be described in reference to **636**, the delay value in delay unit **66** can be updated by adding any necessary delay resulting from the addition of non-adaptive filters to the existing delay value in delay **66**. Flow then proceeds to block **638** where the filter shortening estimation option is disabled. That is, the flow of

Note that in the embodiment illustrated in **630**-**638** are performed serially after determining that the ERLE is good enough. However, in alternate embodiments, blocks **630**-**638** (or a subset of blocks **630**-**638**) can be launched as a separate thread performed in parallel to other tasks of echo canceller **20**. Upon completion of the flow of **20** that the results are ready such that adaptive filter **62** can be updated. Alternatively, echo canceller **20** can be alerted that the entire filter shortening has been completed. For example, a signal or interrupt may be provided to echo canceller **20** to indicate completion of blocks **630**-**638** (or a subset of blocks **630**-**638**).

**630** of **62**. Flow begins with block **640** where the magnitude of the coefficients of adaptive filter **62** are moved to a circular buffer. That is, a snapshot of the filter coefficients of adaptive filter **62** is taken and stored in a circular buffer of size N, having locations 0 to N−1. The current coefficients of adaptive filter **62**, H, can be represented as H=[h_{0}, . . . , h_{N−1}], where h_{0}, . . . , h_{N−1 }correspond to the coefficients and N corresponds to the number of coefficients or taps of adaptive filter **62**. Therefore, the values corresponding to |h_{n}| are stored in a circular buffer at location “n MOD N”, where n corresponds to the sample number, and “|x|” indicates the “magnitude of x” (and corresponds to positive values). The expression “n MOD N” corresponds to the modulus of n which refers to the remainder of the operation n/N. For example, if N is 256 and the value of n is 270, n MOD N refers to **14**, where the value of |h_{270}| is wrapped around from “location 270” (which is beyond the range of the circular buffer of size N) to location **14** of the circular buffer. Therefore, if the value of n is greater than N, the value of N can be continuously subtracted from n until n falls within the range of 0 to N−1. Similarly, if the value of n is less than 0, then the value of N can be continuously added to n until n falls within the range of 0 to N−1.

Flow then proceeds to block **642** where for every coefficient, h, the energy E(n) (for n=0 to N−1) is computed as the sum of the magnitude values within a sliding window of size LW. In one embodiment, LW is related to the length of the target window size, i.e. the target number of taps or coefficients after reducing the effective number of coefficients of adaptive filter **62**. For example, in one embodiment, LW may correspond to a sliding window of size 10 samples, where 10 taps is the desired filter length. Therefore, if N is 256 (indicating 256 coefficients h of adaptive filter **62**), then 256 values of E(n) are determined where each of the 256 values of E(n) is a sum of 10 magnitudes (corresponding to the 10 samples within LW). E(n) can therefore be expressed as shown below in equation 40.

In the above equation (and other equations described herein), note that the notation [X]_{N }corresponds to X MOD N. Flow then proceeds from block **642** to block **644** where the delay, D, is set as the location of the energy peak minus LW. That is, after N values of E (the energy within the sliding window LW) are determined, the point (sample time) at which the maximum E occurs minus LW corresponds to the pure delay of the impulse response. Therefore, D can be expressed as shown in equation 41.

*D*=arg (max *E*(*n*))−*LW *for *n=*0, 1, . . . , *N−*1 Equation 41:

In the above equation, max E(n) refers to the maximum value of E of the N values of E taken, and arg(max E(n)) refers to the argument or point at which the maximum E occurs, where the “argument” corresponds to the sample time. Therefore, D corresponds to the pure delay.

Flow then proceeds to block **646** where the dispersion time corresponds to the number of samples between D and the next location where the energy, E, is smaller than a predetermined threshold. That is, the general trend of E(n) (corresponding to the magnitude of the impulse response) over the range of n=0, 1, . . . , N−1 can be described as generally increasing to a maximum peak, and then decreasing back down. Therefore, after reaching the maximum value, E(n) decreases, and the point at which it decreases beyond a predetermined threshold corresponds to the end of the dispersion time, where the dispersion time is therefore the number of samples between D and the point (sample time) at which E(n) reaches the predetermined threshold after having achieved its peak value. The predetermined value can be set to any value which can indicate the end of the dispersion time. For example, in one embodiment, it may be set to 192 samples (i.e. 24 milliseconds at an 8 KHz sampling rate).

**632** of **64** or non-adaptive filters **31** and **35**) and (2) the shortened version of the coefficients of adaptive filter **62** are determined. Flow begins with block **648** where the new filter coefficients W (for either non-adaptive filter **64** or non-adaptive filters **31** and **35**) are determined (note that the details of block **648** will be described in reference to **650** where the coefficients W are convolved with the current adaptive filter **62** coefficients (i.e. with the snapshot taken in block **640** of **62** after the addition of non-adaptive filters **64** or **31** and **35**. Flow proceeds to block **652** where a new pure delay, D, and dispersion time of the shortened filter coefficients B are determined. Therefore, the flow of **40** and **41** may be used to accomplish block **652**. Flow then proceeds to block **654** where the new adaptive filter coefficients for adaptive filter **62** are determined from B (i.e. a portion of B with predefined length), and the maximum number of filter coefficients to be adapted is selected (i.e. number of samples in the selected portion of B).

**636** of **62** is reconfigured. Flow begins with block **656** where the current adaptive filter coefficients, H, are replaced with the shortened coefficients determined previously in block **654** of **658** where the new coefficients W (determined in block **648** of **64** or filters **31** and **35**). Therefore, by loading W into the non-adaptive filter or filters, they are enabled to allow adaptive filter **62** to have a reduced filter length. Flow then proceeds to block **660** where delay unit **66** within adaptive filter **28** is updated with a new delay. For example, one embodiment may simply update delay unit **66** with the newly determined pure delay, D. An alternate embodiment may determine the delay currently stored in delay unit **66** and update the existing value as necessary. Alternatively, if the new delay does not vary much from the existing delay within delay unit **66**, delay unit **66** may not be updated at all. Also, in block **660**, once the adaptive filter **62** is loaded with the new coefficients in block **656**, it must be configured to adapt the new number of coefficients.

**648** of **64** or filters **31** and **35**). Flow begins at decision diamond **662** where it is determined whether any pre-computed filter coefficients W exist. For example, a library including a variety of different possible sets of W corresponding to different hybrid and channel conditions may exist which have been precomputed, and therefore a new W can simply be selected from the library and flow would proceed to block **650** of **664** where any method to determine new filter candidates may be used to find W. Various embodiments for determining W will be described in more detail in reference to

Flow then proceeds to block **666** where the roots of W are determined. For example, W can be expressed as W=[w_{0}, w_{1}, . . . , w_{M−1}] where w_{0}, w_{1}, . . . , w_{M−1 }correspond to the filter coefficients and M corresponds to the number of filter coefficients such that W(z) can be expressed as shown below in equation 42:

*W*(*z*)=(*w* _{0} *z* ^{M−1} *+w* _{1} *z* ^{M−2} *+ . . . +w* _{M−2} *z+w* _{M−1})*z* ^{1−M} Equation 42:

To determine the roots of W, W(z) is set to 0 and solved for z, where z has M−1 solutions. Therefore, the roots R of W(z) can be expressed as R=[r_{0}, r_{1}, . . . , r_{M−1}] where the roots include complex numbers and their conjugates. W(z)=0 can therefore also be expressed as shown below in equation 43:

*W*(*z*)=0=(*z−r* _{0})(*z−r* _{1}) . . . (*z−r* _{M−2})(*z−r* _{M−1})*z* ^{1−M} Equation 43:

Flow then proceeds to block **668** where additional constraints to the roots are imposed such that W can be used in either FIR or IIR processing mode and remain stable. For example, if W is used in an FIR implementation, no imposition of constraints is necessary to ensure stability, but if in an IIR implementation, the roots must be within the unit circle. The unit circle refers to a circle in a plane defined by an x-axis corresponding to real numbers and a y-axis corresponding to imaginary numbers. The unit circle is drawn about the origin (the intersection of the x- and y-axis, corresponding to 0 on both the x- and y-axes), and has a radius of 1. For each root r_{k }of W(z) that does not lie within the unit circle, r_{k }is transposed such that it does lie within the unit circle. Alternatively, rather than imposing that the roots be within the unit circle, the constraints can be imposed such that the roots of W(z) be within a circle centered about the origin having a radius of ρ, where ρ is less than 1. Therefore, in this embodiment, for each root r_{k }of W(z) that does not lie within the circle having radius ρ, the root r_{k }is transposed such that it does lie within the circle having radius ρ. Therefore, if |r_{k}| (i.e. the magnitude of r_{k}) is greater than ρ, the transposition can be expressed as shown in equation 44:

Note that in the above equation r_{k}* denotes the complex conjugate of r_{k}. Flow then proceeds to block **670** where the new filter coefficients W_{new }are constructed from the modified roots. That is, if any of the roots had to be modified due to the constraints imposed in block **668**, W_{new }is determined using the modified roots. The new roots can be substituted into equation 43 above for r_{0}, . . . , r_{M−1 }to provide a new W(z) (i.e. W_{new}). W_{new }is then used as the filter coefficients W for the remainder of the flow (which continues with block **650** of

**704** where, for all design methods, a solution W is determined for every channel impulse response in the training set. Therefore, if 2 design methods are used for 8 channel impulse responses in the training set, a total of 16 solutions (W_{0}-W_{15}) are determined. Furthermore, the method of **706** where, for every solution W, a convolution, B_{k}, of W and every impulse response in the training set is estimated. Therefore, in the current example where there are 16 solutions (W_{0}-W_{15}) and 8 impulse responses in the training set, a total of 128 convolutions are estimated. For any solution W, each B_{k }can therefore be expressed as B_{k}=[b_{k,0},b_{k,1}, . . . , b_{k,N−1}] where b_{k,0},b_{k,1}, . . . , b_{k,N−1 }are the coefficients and N is the length of B_{k}. Note that [k]_{8 }indicates the channel number in the training set.

Flow then proceeds to block **708** where for every B_{k}, the dispersion region of desirable length (i.e. target filter length) having the maximum energy is located. That is, the dispersion region of maximum energy can be located using equations 40 and 41 above, where LW corresponds to the desirable length of the dispersion region. Therefore, the equations for block **708** can be expressed as follows.

The above equation is analogous to equation 40 described above. In equation 45, N is the length of the particular B_{k }and E_{k}(n) corresponds to the energy within a sliding window LW (corresponding to the desirable length of the dispersion regions). Therefore, N energy values (E_{k}(0), . . . , E_{k}(N−1)) are determined for each B_{k}. The dispersion region having maximum energy for each B_{k }can therefore be determined using the following equation.

*D* _{k} *=arg*(max *E* _{k}(*n*))−*LW *for *n=*0, 1, *. . . , N−*1 Equation 46:

Equation 46 is analogous to equation 41 described above. In equation 46, D_{k }corresponds to the pure delay of B_{k }at the maximum energy, and the dispersion region is therefore the region beginning with D_{k}, ending with D_{k}+LW, and having maximum energy. (Alternatively, the ending of the dispersion region can be defined as the point at which the energy E_{k}(n) falls below a predetermined threshold, as was described above in reference to equation 41.) Therefore, for each B_{k}, a dispersion region of maximum energy is determined. Flow then proceeds to block **710** where, for every B_{k}, a figure of merit, FM_{k}, is estimated. The figure of merit is defined as the ratio of maximum energy, max E_{k}(n) (from block **708**) and the total energy, E_{k}, of B_{k}. Therefore, E_{k }can be determined using the following equations:

In the above equation, N refers to the length of B_{k}. The figure of merit can therefore be expressed as follows.

In the above equation, N refers to the length of B_{k}. Flow proceeds to block **712** where an average figure of merit FM_{AVG }for all channel impulse responses is determined. Every solution W_{k }(k=0,1, . . . 15, in the above example) will have its own average figure of merit FM_{AVG}. Flow proceeds to block **714** where the optimal filter W is selected such that FM_{AVG }is maximized among all possible design methods. The method of **20** and loaded into non-adaptive filters **64**, **31**, or **35**, as needed.

Note that although the above descriptions of W assume the additional filters are non-adaptive, adaptive filters may be used in place of non-adaptive filters **64**, **31**, and **35**. However, constraints may need to be imposed on W, as described above to ensure stability, and if the filters are adaptive, then the stability constraints may need to be imposed on a per-sample basis, since the filter adaptation may result in an unstable filter.

**704** of **20** during its operation. In one embodiment, echo canceller **20** may perform all three design methods to determine W (in block **664** of **632** of **668** of **20**, depending on the embodiment. Alternatively, a single design method may be used by echo canceller **20**.

**664** of **672** where the dispersion region of the adaptive filter **62** coefficients are moved to a circular buffer such that the delay, D, (computed in block **644**) is compensated. Therefore, the delay compensated coefficients G can be expressed as G=[g_{0}, g_{1}, . . . , g_{N−1}] where g_{0}, g_{1}, . . . , g_{N−1 }are the coefficients of the delay compensated coefficients G and N is the length of adaptive filter **62**. Therefore, the relationship between H (the uncompensated coefficients of adaptive filter **62** corresponding to the original snapshot taken in block **640** of

*g* _{i} *=h* _{[D−i]} _{ N }for *i=*0, 1, *. . . , N− 1 * Equation 49:

In equation 49, g_{i }are the delay compensated coefficients which are stored in the ith location of the corresponding circular buffer. **1** of **4**+T**2** in **674** where the desirable filter length is defined. For example, in one embodiment, as described above, the desirable filter length is defined to be 10 (where the dispersion time of the impulse response is desired to be compacted into 10 samples). Referring to the example of **5**, the time between 0 and S1, where 0 defines the beginning of the dispersion time (and also corresponds to the first coefficient of G, since G has been delay compensated) and S1 defines the end of the desirable filter length.

Once the desirable filter length is determined, the coefficients of the dispersion region within the desirable filter length are cleared to define the residual coefficients, V. That is, g_{i }is set to 0 for i=0, 1, . . . , S1. Therefore, V can be expressed as V=[0, . . . , 0, v_{0}, v_{1}, . . . , v_{K−1}] where K is the number of non-zero components of V and each non-zero coefficient of V is defined as follows.

*v* _{J} *=g* _{S1+j }for *j=*0, 1, *. . . , K−*1 Equation 50:

Therefore, referring to the example of **5** are set to zero, and the coefficients of V represent the residual distortion, i.e. the portion of the impulse response corresponding to time T**6**. Flow proceeds to blocks **676**-**680** which operate to equalize the residual distortion, V.

In block **676**, the fast Fourier transform (FFT) of V is computed. Flow proceeds to block **678** where the inverse, I, of the FFT(V) is computed, where I=1/FFT(V). Flow proceeds to block **680** where W_{1 }is computed as the inverse FFT (IFFT) of I, where W_{1}=IFFT(I). Therefore, W_{1 }is the inverse of V and can be used to equalize the residual distortion, V. Flow proceeds to block **682** where W is determined from a window of W_{1 }with a pre-defined length having the maximum energy (similarly to the estimation of channel dispersion on **684** where the filter coefficients W are normalized, for example, by the Euclidean magnitude of W (i.e. L_{2 }norm).

**686** where, using the delay compensated filter coefficients G (from block **672** and equation 49), a desirable filter length, as described above, is defined and a convolution matrix C is determined. Convolution matrix C can be defined as follows.

*C=S* _{L} *·C*on*G* Equation 51:

S_{L }corresponds to the selection matrix which may be expressed as follows.

*S* _{L}=[0

In equation 52, 0 is a (N+M−L−1)×L zero matrix and I is an (N+M−L−1)×(N+M−L−1) identity matrix. In this equation, N corresponds to the length of G, M to the length of W (the predefined number of non-adaptive filter taps to accomplish the desirable adaptive filter length), and L to the desirable adaptive filter length. ConG corresponds to the convolution matrix of G and can be expressed as follows.

The matrix C is therefore obtained from the convolution matrix ConG of the delay compensated filter coefficients G by ignoring its initial L rows, which will define the coefficients of the shortened channel by the end of this design process.

Flow proceeds to block **688** where b_{w }is defined as the first row of matrix C of equation 51 placed as a column vector. Flow proceeds to block **690** where a matrix A is computed.

A=C^{T}C Equation 54:

In equation 54, the notation C^{T }refers to the transposition of matrix C. Flow proceeds to block **692** where the system of equations given by AW=b_{w }is solved for W. Flow then proceeds to block **694** where the filter coefficients W are normalized. Anyone skilled in the art will immediately identify that the above solution corresponds to the minimum mean squared error solution of the system CW=[1,0, . . . 0]^{T}, which attempts to equalize distortion by considering the overall convolution of the adaptive filter **62** coefficients with W.

**696** where, using the delay compensated filter coefficients G (from block **672** and equation 49), a desirable filter length, as described above, is defined and a convolution matrix C is determined. Therefore, the same C matrix as defined in Equation 51 is used in **698**, where, as in block **690** of **700** where the maximum solution of W^{T}W/W^{T}AW is estimated. Note that W^{T}W/W^{T}AW provides a ratio of energy of the normalized taps of W weighted by the matrix A, therefore, the maximum solution of W^{T}W/W^{T}AW minimizes the energy W^{T}AW conditioned to W^{T}W=1. Note also that W^{T}W/W^{T}AW is equivalent to W^{T}IW/W^{T}AW where I is the identity matrix, such that the solution W is the generalized eigenvector corresponding to the largest eigenvalue of the pair (I, A), which can be computed using any off-the-shelf algorithm for estimating eigenvectors. Flow then proceeds to block **702** where the filter coefficients W determined in block **700** are normalized.

Therefore, **62** coefficients with W. However, **62** and W. As discussed above, all methods may be implemented by echo canceller **20** and the best solution is chosen either prior to modifying the roots of W or after reconstructing the new filter coefficients W from the modified roots.

In the foregoing specification, the invention has been described with reference to specific embodiments. However, one of ordinary skill in the art appreciates that various modifications and changes can be made without departing from the scope of the present invention as set forth in the claims below. For example, any of the methods taught herein may be embodied as software on one or more of computer hard disks, floppy disks, 3.5″ disks, computer storage tapes, magnetic drums, static random access memory (SRAM) cells, dynamic random access memory (DRAM) cells, electrically erasable (EEPROM, EPROM, flash) cells, nonvolatile cells, ferroelectric or ferromagnetic memory, compact disks (CDs), laser disks, optical disks, and any like computer readable media. Also, the block diagrams may different blocks than those illustrated and may have more or less blocks or be arranged differently. Also, the flow diagrams may also be arranged differently, include more or less steps, be arranged differently, or may have steps that can be separated into multiple steps or steps that can be performed simultaneously with one another. Accordingly, the specification and figures are to be regarded in an illustrative rather than a restrictive sense, and all such modifications are intended to be included within the scope of present invention.

Benefits, other advantages, and solutions to problems have been described above with regard to specific embodiments. However, the benefits, advantages, solutions to problems, and any element(s) that may cause any benefit, advantage, or solution to occur or become more pronounced are not to be construed as a critical, required, or essential feature or element of any or all the claims. As used herein, the terms “comprises,” “comprising,” or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus.

Citations de brevets

Brevet cité | Date de dépôt | Date de publication | Déposant | Titre |
---|---|---|---|---|

US4363100 * | 28 oct. 1980 | 7 déc. 1982 | Northern Telecom Limited | Detection of tones in sampled signals |

US4645883 | 9 mai 1984 | 24 févr. 1987 | Communications Satellite Corporation | Double talk and line noise detector for a echo canceller |

US4658420 | 30 nov. 1984 | 14 avr. 1987 | Nec Corporation | Tone responsible disabler circuit for an echo canceller |

US4751730 | 28 avr. 1986 | 14 juin 1988 | International Business Machines Corp. | Process and system for improving echo cancellation within a transmission network |

US4771396 | 14 mars 1985 | 13 sept. 1988 | British Telecommunications Plc | Digital filters |

US4829566 | 16 avr. 1987 | 9 mai 1989 | Telecommunications Radioelectriques Et Telephoniques T.R.T. | Apparatus for detecting and discriminating phase jumps in a periodic signal and its application to a telephone tone signal with phase inversions |

US4894820 | 21 mars 1988 | 16 janv. 1990 | Oki Electric Industry Co., Ltd. | Double-talk detection in an echo canceller |

US5029204 | 26 oct. 1989 | 2 juil. 1991 | Dsc Communications Corporation | Operational status controller for echo canceling |

US5042026 | 29 févr. 1988 | 20 août 1991 | Nec Corporation | Circuit for cancelling whole or part of a waveform using nonrecursive and recursive filters |

US5164962 | 9 mai 1991 | 17 nov. 1992 | Oki Electric Industry Co., Ltd. | Adaptive equalizer with midburst correction capability |

US5164989 | 11 déc. 1990 | 17 nov. 1992 | Octel Communications Corporation | Echo cancellation methods and apparatus for voice processing systems |

US5274705 | 24 sept. 1991 | 28 déc. 1993 | Tellabs Inc. | Nonlinear processor for an echo canceller and method |

US5283784 | 30 juil. 1992 | 1 févr. 1994 | Coherent Communications Systems | Echo canceller processing techniques and processing |

US5343522 | 4 juin 1992 | 30 août 1994 | Northern Telecom Limited | Adaptive sparse echo canceller using a sub-rate filter for active tap selection |

US5353346 | 22 déc. 1992 | 4 oct. 1994 | Mpr Teltech, Limited | Multi-frequency signal detector and classifier |

US5390250 | 17 déc. 1991 | 14 févr. 1995 | U.S. Phillips Corporation | Echo canceller with improved doubletalk detection |

US5392347 | 26 janv. 1993 | 21 févr. 1995 | Nec Corporation | Ringing tone signal detecting circuit |

US5420921 | 28 oct. 1992 | 30 mai 1995 | Nokia Telecommunications Oy | Method for the detection of a disable tone signal of an echo canceller |

US5446787 | 8 avr. 1993 | 29 août 1995 | Telefonaktiebolaget Lm Ericsson | Method for avoiding self-oscillation in conjunction with echo cancellation |

US5485522 | 29 sept. 1993 | 16 janv. 1996 | Ericsson Ge Mobile Communications, Inc. | System for adaptively reducing noise in speech signals |

US5521908 | 20 avr. 1995 | 28 mai 1996 | Tellabs Operations Inc. | Method and apparatus for providing reduced complexity echo cancellation in a multicarrier communication system |

US5561668 | 6 juil. 1995 | 1 oct. 1996 | Coherent Communications Systems Corp. | Echo canceler with subband attenuation and noise injection control |

US5587996 | 18 sept. 1995 | 24 déc. 1996 | Fujitsu Limited | Method of radio-line relief and radio equipment in SDH network |

US5615302 | 30 sept. 1992 | 25 mars 1997 | Mceachern; Robert H. | Filter bank determination of discrete tone frequencies |

US5631899 | 31 mai 1995 | 20 mai 1997 | Lucent Technologies Inc. | Acoustic echo canceler |

US5646991 | 23 févr. 1996 | 8 juil. 1997 | Qualcomm Incorporated | Noise replacement system and method in an echo canceller |

US5664011 | 25 août 1995 | 2 sept. 1997 | Lucent Technologies Inc. | Echo canceller with adaptive and non-adaptive filters |

US5687229 | 23 févr. 1996 | 11 nov. 1997 | Qualcomm Incorporated | Method for controlling echo canceling in an echo canceller |

US5689556 | 15 sept. 1995 | 18 nov. 1997 | Hughes Electronics | Method of detecting narrow-band signals in an echo canceller |

US5721782 | 25 mars 1996 | 24 févr. 1998 | Motorola, Inc. | Partitioned echo canceler utilizing decimation echo location |

US5737410 | 21 déc. 1994 | 7 avr. 1998 | Nokia Telecommunication Oy | Method for determining the location of echo in an echo canceller |

US5815568 | 31 janv. 1996 | 29 sept. 1998 | Telefoanktiebolaget Lm Ericsson | Disabling tone detector for network echo canceller |

US5818929 * | 13 déc. 1994 | 6 oct. 1998 | Canon Kabushiki Kaisha | Method and apparatus for DTMF detection |

US5877653 * | 18 nov. 1996 | 2 mars 1999 | Samsung Electronics Co., Ltd. | Linear power amplifier and method for removing intermodulation distortion with predistortion system and feed forward system |

US5920548 | 1 oct. 1996 | 6 juil. 1999 | Telefonaktiebolaget L M Ericsson | Echo path delay estimation |

US5920834 | 31 janv. 1997 | 6 juil. 1999 | Qualcomm Incorporated | Echo canceller with talk state determination to control speech processor functional elements in a digital telephone system |

US5949888 | 15 sept. 1995 | 7 sept. 1999 | Hughes Electronics Corporaton | Comfort noise generator for echo cancelers |

US5978473 | 27 déc. 1995 | 2 nov. 1999 | Ericsson Inc. | Gauging convergence of adaptive filters |

US6006083 * | 11 sept. 1997 | 21 déc. 1999 | Nortel Networks Corporation | Tone detection |

US6044068 | 1 oct. 1996 | 28 mars 2000 | Telefonaktiebolaget Lm Ericsson | Silence-improved echo canceller |

US6055310 | 17 déc. 1997 | 25 avr. 2000 | Nortel Networks Corporation | Phase reversal tone detector using DSP |

US6163608 | 9 janv. 1998 | 19 déc. 2000 | Ericsson Inc. | Methods and apparatus for providing comfort noise in communications systems |

US6163609 | 16 avr. 1998 | 19 déc. 2000 | Nokia Telecommunications Oy | System and method for echo cancelling and mobile communications device |

US6185195 | 16 mai 1997 | 6 févr. 2001 | Qualcomm Incorporated | Methods for preventing and detecting message collisions in a half-duplex communication system |

US6195430 | 17 juin 1998 | 27 févr. 2001 | Telefonaktiebolaget Lm Ericsson | Method and device for echo cancellation using power estimation in a residual signal |

US6263078 | 7 janv. 1999 | 17 juil. 2001 | Signalworks, Inc. | Acoustic echo canceller with fast volume control compensation |

US6266367 | 28 mai 1998 | 24 juil. 2001 | 3Com Corporation | Combined echo canceller and time domain equalizer |

US6282286 | 31 août 1998 | 28 août 2001 | Mitel Corporation | Nonlinear processor for acoustic echo canceller with background noise preservation and long echo tail suppression |

US6321200 | 2 juil. 1999 | 20 nov. 2001 | Mitsubish Electric Research Laboratories, Inc | Method for extracting features from a mixture of signals |

US6415139 * | 24 nov. 1998 | 2 juil. 2002 | Oki Electric Industry Co, Ltd. | Detection circuit of tone signal |

US6563803 | 24 nov. 1998 | 13 mai 2003 | Qualcomm Incorporated | Acoustic echo canceller |

US6654463 | 28 mai 1999 | 25 nov. 2003 | 3Com Corporation | Round trip delay estimator and compensator for the echo canceller |

US6738358 | 6 sept. 2001 | 18 mai 2004 | Intel Corporation | Network echo canceller for integrated telecommunications processing |

US6768796 | 5 févr. 2001 | 27 juil. 2004 | 3Com Corporation | System and method for echo cancellation |

US6914979 * | 25 avr. 2001 | 5 juil. 2005 | Adaptive Digital Technologies, Inc. | Tone detection |

US20020039415 | 28 sept. 2001 | 4 avr. 2002 | Dieter Schulz | Noise level calculator for echo canceller |

US20020154760 * | 20 août 2001 | 24 oct. 2002 | Scott Branden | Tone relay |

US20020181414 | 6 avr. 2001 | 5 déc. 2002 | Yhean-Sen Lai | Method and apparatus for detecting robbed bit location in PCM modems and the like |

US20030133565 | 15 janv. 2002 | 17 juil. 2003 | Chienchung Chang | Echo cancellation system method and apparatus |

US20030185292 | 2 avr. 2002 | 2 oct. 2003 | Fernandez-Corbaton Ivan Jesus | Adaptive filtering with DC bias compensation |

US20050195967 * | 8 mars 2004 | 8 sept. 2005 | Pessoa Lucio F.C. | Tone event detector and method therefor |

EP0300265A2 | 5 juil. 1988 | 25 janv. 1989 | Advanced Micro Devices, Inc. | Digital tone detection method |

EP0792029A2 | 18 févr. 1997 | 27 août 1997 | Lucent Technologies Inc. | Echo canceller E-side speech detector |

WO1995017784A1 | 21 déc. 1994 | 29 juin 1995 | Nokia Telecommunications Oy | Method for determining the location of echo in an echo cancellar |

WO1997049196A2 | 18 juin 1997 | 24 déc. 1997 | Nokia Telecommunications Oy | Echo suppressor and non-linear processor of echo canceller |

WO2000030325A1 | 30 oct. 1998 | 25 mai 2000 | Nortel Networks Limited | Method and apparatus for detecting signalling tones |

WO2000069080A2 | 19 avr. 2000 | 16 nov. 2000 | Telefonaktiebolaget Lm Ericsson | Pure delay estimation |

Citations hors brevets

Référencé par

Brevet citant | Date de dépôt | Date de publication | Déposant | Titre |
---|---|---|---|---|

US8374851 * | 30 juil. 2007 | 12 févr. 2013 | Texas Instruments Incorporated | Voice activity detector and method |

US9172816 | 30 août 2013 | 27 oct. 2015 | Microsoft Technology Licensing, Llc | Echo suppression |

US9277059 | 28 août 2013 | 1 mars 2016 | Microsoft Technology Licensing, Llc | Echo removal |

US9467571 | 28 août 2013 | 11 oct. 2016 | Microsoft Technology Licensing, Llc | Echo removal |

US9521264 | 28 août 2013 | 13 déc. 2016 | Microsoft Technology Licensing, Llc | Echo removal |

US20080273481 * | 1 mai 2008 | 6 nov. 2008 | Mediaphy Corporation | Warm start receiver |

US20090036170 * | 30 juil. 2007 | 5 févr. 2009 | Texas Instruments Incorporated | Voice activity detector and method |

US20110013766 * | 15 juil. 2009 | 20 janv. 2011 | Dyba Roman A | Method and apparatus having echo cancellation and tone detection for a voice/tone composite signal |

Classifications

Classification aux États-Unis | 379/386 |

Classification internationale | H04M1/60, H04M3/00, H04B3/23 |

Classification coopérative | H04B3/23 |

Classification européenne | H04B3/23 |

Événements juridiques

Date | Code | Événement | Description |
---|---|---|---|

24 juin 2002 | AS | Assignment | Owner name: MOTOROLA, INC., ILLINOIS Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNORS:PESSOA, LUCIO F. C.;DYBA, ROMAN A.;HE, PERRY P.;REEL/FRAME:013059/0500 Effective date: 20020621 |

7 mai 2004 | AS | Assignment | Owner name: FREESCALE SEMICONDUCTOR, INC., TEXAS Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:MOTOROLA, INC;REEL/FRAME:015360/0718 Effective date: 20040404 Owner name: FREESCALE SEMICONDUCTOR, INC.,TEXAS Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:MOTOROLA, INC;REEL/FRAME:015360/0718 Effective date: 20040404 |

24 sept. 2008 | AS | Assignment | Owner name: CITIBANK, N.A., NEW YORK Free format text: SECURITY AGREEMENT;ASSIGNOR:FREESCALE SEMICONDUCTOR, INC.;REEL/FRAME:021570/0449 Effective date: 20080728 Owner name: CITIBANK, N.A.,NEW YORK Free format text: SECURITY AGREEMENT;ASSIGNOR:FREESCALE SEMICONDUCTOR, INC.;REEL/FRAME:021570/0449 Effective date: 20080728 |

13 mai 2010 | AS | Assignment | Owner name: CITIBANK, N.A., AS COLLATERAL AGENT,NEW YORK Free format text: SECURITY AGREEMENT;ASSIGNOR:FREESCALE SEMICONDUCTOR, INC.;REEL/FRAME:024397/0001 Effective date: 20100413 Owner name: CITIBANK, N.A., AS COLLATERAL AGENT, NEW YORK Free format text: SECURITY AGREEMENT;ASSIGNOR:FREESCALE SEMICONDUCTOR, INC.;REEL/FRAME:024397/0001 Effective date: 20100413 |

12 avr. 2011 | CC | Certificate of correction | |

23 sept. 2011 | FPAY | Fee payment | Year of fee payment: 4 |

18 juin 2013 | AS | Assignment | Owner name: CITIBANK, N.A., AS NOTES COLLATERAL AGENT, NEW YOR Free format text: SECURITY AGREEMENT;ASSIGNOR:FREESCALE SEMICONDUCTOR, INC.;REEL/FRAME:030633/0424 Effective date: 20130521 |

6 nov. 2013 | AS | Assignment | Owner name: CITIBANK, N.A., AS NOTES COLLATERAL AGENT, NEW YOR Free format text: SECURITY AGREEMENT;ASSIGNOR:FREESCALE SEMICONDUCTOR, INC.;REEL/FRAME:031591/0266 Effective date: 20131101 |

17 déc. 2015 | FPAY | Fee payment | Year of fee payment: 8 |

21 déc. 2015 | AS | Assignment | Owner name: FREESCALE SEMICONDUCTOR, INC., TEXAS Free format text: PATENT RELEASE;ASSIGNOR:CITIBANK, N.A., AS COLLATERAL AGENT;REEL/FRAME:037354/0719 Effective date: 20151207 Owner name: FREESCALE SEMICONDUCTOR, INC., TEXAS Free format text: PATENT RELEASE;ASSIGNOR:CITIBANK, N.A., AS COLLATERAL AGENT;REEL/FRAME:037356/0553 Effective date: 20151207 Owner name: FREESCALE SEMICONDUCTOR, INC., TEXAS Free format text: PATENT RELEASE;ASSIGNOR:CITIBANK, N.A., AS COLLATERAL AGENT;REEL/FRAME:037356/0143 Effective date: 20151207 |

12 janv. 2016 | AS | Assignment | Owner name: MORGAN STANLEY SENIOR FUNDING, INC., MARYLAND Free format text: ASSIGNMENT AND ASSUMPTION OF SECURITY INTEREST IN PATENTS;ASSIGNOR:CITIBANK, N.A.;REEL/FRAME:037486/0517 Effective date: 20151207 |

13 janv. 2016 | AS | Assignment | Owner name: MORGAN STANLEY SENIOR FUNDING, INC., MARYLAND Free format text: ASSIGNMENT AND ASSUMPTION OF SECURITY INTEREST IN PATENTS;ASSIGNOR:CITIBANK, N.A.;REEL/FRAME:037518/0292 Effective date: 20151207 |

16 juin 2016 | AS | Assignment | Owner name: MORGAN STANLEY SENIOR FUNDING, INC., MARYLAND Free format text: SUPPLEMENT TO THE SECURITY AGREEMENT;ASSIGNOR:FREESCALE SEMICONDUCTOR, INC.;REEL/FRAME:039138/0001 Effective date: 20160525 |

8 nov. 2016 | AS | Assignment | Owner name: NXP USA, INC., TEXAS Free format text: CHANGE OF NAME;ASSIGNOR:FREESCALE SEMICONDUCTOR INC.;REEL/FRAME:040652/0180 Effective date: 20161107 |

Faire pivoter